JP5960029B2 - Abnormal cause identification system - Google Patents

Description

  The present invention relates to an abnormality cause identification system.

  In a large-scale apparatus such as a CFB (circulating fluidized bed) boiler, when some abnormality occurs, a plurality of types of alarms may be generated in a chain. When multiple types of alarms are generated, it is difficult to specify which alarm is associated with a true abnormality. As a method for identifying the cause of an abnormality in such a large-scale apparatus, for example, as described in Patent Document 1, a priority is set in advance for an alarm (alarm) when an abnormality occurs, and an alarm with a high priority is set. Is known as a representative alarm and other alarms are suppressed to identify a cause of abnormality based on an alarm having a high priority.

Japanese Patent Laid-Open No. 9-16875

  Here, in the above-described abnormality cause identification system, as described above, the abnormality cause is identified based only on the preset priority order of alarms. However, there is a possibility that the cause of the abnormality cannot be accurately identified because the cause related to the alarm having a high priority order is not necessarily the cause of the true abnormality.

  Then, this invention makes it a subject to provide the abnormality cause identification system which can pinpoint an abnormality cause with a sufficient precision.

  In order to solve the above problems, an abnormality cause identification system according to the present invention is an abnormality cause identification system that identifies an abnormality cause of a device that issues a plurality of types of alarms, and a statistical model relating to an occurrence probability of an abnormality cause corresponding to an alarm Includes a model storage unit that is stored in advance, an input alarm information, an inference unit that infers the cause of abnormality based on a statistical model, and a display unit that displays information on the cause of abnormality inferred by the inference unit, The inference unit obtains the abnormal cause candidate based on the alarm information and the statistical model for each predetermined time width divided for a certain period of time, calculates the frequency for each type of the abnormal cause candidate, and at least the type of the type with the highest frequency Inferring an abnormal cause candidate as an abnormal cause.

  In this abnormality cause identification system, an abnormality cause candidate can be obtained from the probability of occurrence of an abnormality cause in a predetermined time width unit, and the abnormality cause can be identified from the frequency of the abnormality cause candidate. As a result, even when a plurality of types of alarms are generated, the cause of the abnormality can be accurately identified.

  The statistical model is preferably created based on alarm observation data corresponding to the cause of the abnormality. In this case, a statistical model can be created from the observation data.

  Moreover, it is preferable that a display part further displays the information regarding the countermeasure of an abnormality cause. In this case, it is possible not only to specify the cause of the abnormality but also to display a specific countermeasure, and it is possible to converge the abnormal state at an early stage.

  Further, as a configuration that favorably exhibits the above-described effects, specifically, the statistical model is preferably a Bayesian network model. In this case, the cause of the abnormality can be specifically inferred using a Bayesian network model.

  Further, the Bayesian network model has a conditional probability table, and the conditional probability table has a plurality of condition nodes related to alarm information and a plurality of result nodes related to the cause of abnormality, and each value of the plurality of condition nodes. Is a value related to the presence or absence of an alarm, at least the value of each of the plurality of condition nodes, the result node, and the random variable are associated with each other, and the inference unit determines the value of each of the plurality of input condition nodes. Of the result nodes associated with the value of the condition node, it is preferable to derive the result node having the largest random variable as an abnormality cause candidate. In this case, a conditional probability table of condition nodes, result nodes, and random variables can be held as a statistical model. Then, the abnormal cause candidate can be obtained by searching from the conditional probability table.

  Moreover, it is preferable that the said apparatus is a circulating fluidized bed boiler. In general, in a circulating fluidized bed (CFB) boiler, it is particularly difficult to identify the cause of abnormality when multiple types of alarms are generated. Therefore, in the present invention for specifying the cause of abnormality of the circulating fluidized bed boiler, the above-described effect of specifying the cause of abnormality with high accuracy is particularly effective.

  ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to provide the abnormality cause identification system which can pinpoint an abnormality cause with a sufficient precision.

It is a figure which shows the functional block of the abnormality cause identification system which concerns on one Embodiment of this invention. It is a figure which illustrates a Bayesian network model. It is a figure which illustrates a conditional probability table | surface. It is a figure explaining the Bayesian network model used with the abnormality cause identification system shown in FIG. It is a table | surface which shows the example of a tag setting regarding an alarm. It is a table | surface which shows the conditional probability table | surface of the Bayesian network model shown in FIG. It is a flowchart which shows the process of the abnormality cause identification system shown in FIG. It is a table | surface which shows the tag regarding the warning for every time width | variety, and an abnormality cause candidate. It is a graph which shows the frequency of an abnormal cause candidate. It is a flowchart which shows the preparation process of a conditional probability table | surface.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the following description, the same or equivalent elements will be denoted by the same reference numerals, and redundant description will be omitted.

  First, each function of the abnormality cause identification system will be described with reference to FIG. FIG. 1 is a diagram showing functional blocks of an abnormality cause identifying system according to an embodiment of the present invention. As shown in FIG. 1, an abnormality cause identifying system 10 according to the present embodiment is a system that identifies an abnormality cause of a boiler plant that issues multiple types of alarms. The boiler plant here is a CFB boiler, and an alarm is issued when any abnormality occurs in the CFB boiler.

  The abnormality cause identification system 10 includes a model storage unit 11, an alarm input unit 12, an alarm storage unit 13, an abnormality cause inference unit (inference unit) 14, a data output unit 15, and a display unit 16. Configured. The abnormality cause identification system 10 physically includes hardware such as a CPU (Central Processing Unit), a main storage device such as a RAM (Random Access Memory) and a ROM (Read Only Memory), and an auxiliary storage device. By operating these, the functions of the above functional blocks are exhibited.

  The model storage unit 11 stores in advance a statistical model related to the probability of occurrence of an abnormality cause corresponding to an alarm. In the present embodiment, a Bayesian network model is used as the statistical model, and the Bayesian network model has a conditional probability table. First, an overview of the Bayesian network and the conditional probability table will be described below.

  A Bayesian network is a data structure for storing a probabilistic causal relationship between a plurality of random variables. By using a Bayesian network, complex knowledge can be expressed efficiently, and various probabilistic inferences can be made based on the knowledge. A Bayesian network includes nodes representing random variables and links representing causal relationships between random variables. A statistical model based on the structure of a Bayesian network is defined and described as a Bayesian network model in this embodiment.

Such a Bayesian network maintains what is called a conditional probability table (CPT) for each node. The conditional probability refers to the probability that another node B takes a certain value under the condition that the node A takes a certain value, and is expressed as P (B | A). Further, the probability that the node A and the node B occur simultaneously is called “simultaneous probability” and is expressed as P (A∩B). Since the conditional probability represents the probability that the node B will occur when the node A is considered as a whole, it can be expressed by the following equation (1) using the joint probability.

  That is, the conditional probability table is a table showing the conditional probabilities that the node B has a certain value when the set of the node A that is the parent node of the node B has a certain value combination. In the present embodiment, a node corresponding to node A in the above equation is described as a condition node, and a node corresponding to node B is described as a result node.

On the other hand, the probability of occurrence of node A when node B occurs can be expressed by the following equation (2).

Since the above equation (1) and the above equation (2) have a common joint probability P (A∩B), the following equation (3) is obtained.

  Equation (3) above shows that the probability that node B will occur when node A occurs can be calculated using the probability that node A will occur when node B occurs. Such P (A | B) is called an inverse probability of P (B | A).

  FIG. 2 is a diagram illustrating a Bayesian network model, and FIG. 3 is a diagram illustrating a conditional probability table. As shown in FIG. 2, in this example, three types of nodes representing random variables are described (random variables A, B, and X), and the random variable A is “whether an abnormality A has occurred”, the probability The variable B represents “whether an abnormality B has occurred”, and the random variable X represents “whether an alarm has been issued”.

  The causal relationship between the three types of nodes is given as a conditional probability table as shown in FIG. The conditional probability table is data that stores knowledge of strength in the causal relationship between nodes. For example, P (X = reported | A = occurs, B = occurs) = 0.950 indicates that when anomaly A and anomaly B that are condition nodes I occur simultaneously, an alarm that is a result node O It represents the knowledge that the probability of being issued is 0.950. P (A = occurs) represents knowledge that the probability of occurrence of abnormality A is 0.279.

Here, by using the above-described inverse probability, P (A = occurs | X = notified) from the conditional probability table (conditional probability table for P (X | A, B)) shown in FIG. ), P (B = occurs | X = reported). P (A | X) is the probability that the cause when the alarm is issued is abnormal A, and P (B | X) is also the cause when the alarm is issued as B. Represents the probability of being. When transformed based on the Bayesian equation, P (A | X) is obtained as follows.

Similarly, P (B | X) is obtained as follows.

  As shown in the above equation, the probability P (A | X) caused by the abnormality A when the alarm is issued is 0.948, and the probability that the abnormality B is caused when the alarm is issued. When P (B | X) is 0.113, it can be seen that the probability that the abnormality A is the cause is higher when the alarm is issued.

Further, a probability P (A∩B | X) that an abnormality A occurs when an alarm is issued and an abnormality B also occurs is obtained as follows.

  It can be seen that the probability that the abnormality A and the abnormality B occur simultaneously when the alarm is issued is smaller than 0.948 of P (A | X) and 0.113 of P (B | X). As described above, when the Bayesian network is used, not only the probability that a certain result occurs when a certain cause occurs but also the cause can be inferred using the probability that the corresponding result occurs. In the above example, two condition nodes I and one result node O are used. However, the present invention is not limited to this, and the condition node I and the result node O may be one or two or more.

  Returning to FIG. 1, the model storage unit 11 of the present embodiment stores a conditional probability table H (see FIG. 6) created based on the observation data of the alarm corresponding to the cause of the abnormality, as will be described later. Yes. The alarm input unit 12 is connected to a control device (not shown) of the CFB boiler, and receives alarm information from this control device. The alarm information is information regarding a plurality of types of alarms issued in the CFB boiler, and includes at least information regarding the presence or absence of the alarm.

  The alarm storage unit 13 stores the alarm information received by the alarm input unit 12. The alarm accumulation unit 13 accumulates alarm information for a predetermined time by dividing it into a predetermined time width, and specifically, a combination of alarms for a predetermined time width is one data. The accumulated alarm information is used for derivation of abnormality cause candidates by an abnormality cause inference unit 14 described later.

  The abnormality cause inference unit 14 infers the cause of the abnormality based on the input alarm information and the statistical model. The abnormality cause inference unit 14 here reads a statistical model from the model storage unit 11 and infers an abnormality cause by inputting alarm information to the statistical model. Specifically, the abnormality cause inference unit 14 uses a combination of alarms for each time width stored in the alarm storage unit 13 as the alarm information, and each abnormality cause candidate for each predetermined time width divided in a certain time. Is calculated for each type of abnormality cause candidate, and the abnormality cause candidate of the type having the highest frequency is inferred as an abnormality cause.

  Note that the abnormality cause inference unit 14 may infer only the abnormality cause candidate of the type with the highest frequency as the abnormality cause, or rearrange the abnormality cause candidates in descending order of the frequency, It may be inferred as a cause.

  As will be described later, the abnormality cause inference unit 14 of the present embodiment infers an abnormality cause using the conditional probability table H (see FIG. 6). In the conditional probability table H of FIG. 6, the condition node I corresponds to the alarm information, and the result node O corresponds to the cause of the abnormality. Further, the values of the plurality of condition nodes I, the plurality of result nodes O, and the random variables are associated with each other. After determining the value of each of the plurality of condition nodes I from the input alarm information, the abnormality cause inference unit 14 has the largest random variable among the result nodes O associated with the determined value of the condition node I. Is derived as an abnormal cause candidate.

  The data output unit 15 outputs the abnormality cause inferred by the abnormality cause inference unit 14. As described above, the abnormality cause inference unit 14 can infer one abnormality cause as the abnormality cause, and can also infer a plurality of abnormality cause candidates having higher frequencies as the cause of abnormality. 15 outputs one or more abnormality causes according to the inference result of the abnormality cause inference unit 14.

  The display unit 16 is for the operator to visually confirm the driving situation, and displays information regarding the cause of the abnormality output by the data output unit 15 on a monitor screen or the like. The display unit 16 may further display information on countermeasures for the cause of abnormality in addition to information on the cause of abnormality.

  Next, a process related to abnormality cause identification in the abnormality cause identification system 10 will be described.

  FIG. 4 is a diagram for explaining a Bayesian network model used in the anomaly cause identification system, FIG. 5 is a table showing a setting example of a tag related to an alarm, and FIG. 6 is a table showing a conditional probability table of the Bayesian network model. FIG. 7 is a flowchart showing the processing of the abnormality cause identification system. In FIG. 6, a part of the conditional probability table H used in the abnormality cause identification system 10 is omitted.

  As shown in FIG. 4, the Bayesian network model used in the abnormality cause identifying system 10 here has TAG0, TAG1, and TAG2 as the condition nodes I related to alarm information. Further, as a result node O related to the cause of the abnormality, there are an abnormality A, an abnormality B, an abnormality C, and an abnormality D. That is, the Bayesian network model in FIG. 4 represents a model in which a parent node (cause) is a cause of abnormality and a child node (result) is a tag related to an alarm.

  TAG0, TAG1, and TAG2 are assigned tags for each alarm of the CFB boiler in order to facilitate data handling. For example, as shown in FIG. 5, the alarm content and the tag are associated with each other. Similarly, abnormalities A to D are obtained by assigning abnormal numbers to various abnormalities. An example of an abnormality is a power outage.

  Note that the Bayesian network model in FIG. 4 shows a part of the tag relating to the alarm that is the condition node I, and may actually have a plurality of tags relating to the alarm that are not shown. Similarly, the Bayesian network model shown in FIG. 4 shows a part of the cause of abnormality that is the result node O, and may actually have a plurality of causes of abnormality that are not shown. In the link L indicating the causal relationship between the condition node I and the result node O, an arrow is directed from each result node O to each condition node I.

  As shown in FIG. 6, the conditional probability table H here is an abnormal cause (result node O) under the condition that the tags (TAG0, TAG1,. (Abnormality A, abnormality B,...) Represents a probability (probability variable) that is a candidate for an abnormality. The value taken by the tag related to the alarm is a value related to the presence or absence of the alarm and is either with or without the alarm (in FIG. 6, “present” and “absent” are shown).

  In such an abnormality cause identification system 10, first, alarm information issued from the control device of the CFB boiler is received by the alarm input unit 12 (S1). Subsequently, the alarm accumulation unit 13 accumulates alarms divided into predetermined (arbitrary) time widths (S2).

  FIG. 8 is an example showing tags and alarm cause candidates for alarms for each predetermined time width. In the example shown in FIG. 8, the time width is 1 second. As shown in FIG. 8, a tag related to an alarm (alert alarm) received by the alarm input unit 12 at a certain time is 1, and a tag related to an alarm not received (alarm not issued) is 0. Alarm tags are aggregated.

  Subsequently, based on the alarm stored in the alarm storage unit 13 and the conditional probability table H of the Bayesian network model, the abnormality cause inference unit 14 derives one abnormality cause candidate for each predetermined time width (S3). ). Specifically, the value of the condition node I in the conditional probability table H is determined, and the abnormality having the highest probability among the plurality of result nodes O associated with the determined value of the condition node I is the predetermined time width. Is derived as an abnormal cause candidate.

  Here, as shown by a thick frame in FIG. 6, it is assumed that TAG0: existence (alarm is present), TAG1: existence, TAG2: existence, TAG3: existence is determined as the value of each condition node I in a certain time width. (Other tag values are omitted). In this case, out of the plurality of result nodes O associated with the value of the condition node I, the abnormality A having the highest probability of 58% is derived as an abnormality cause candidate in the predetermined time width.

  Subsequently, the abnormality cause inference unit 14 counts the frequency of the abnormality cause candidates for each predetermined time width for a certain period of time for each type of abnormality cause candidate (S4). An example of the result of counting the frequency of abnormal cause candidates is shown in FIG. FIG. 9 is a graph showing the frequency of abnormal cause candidates. In FIG. 9, the horizontal axis is plotted in the frequency distribution with the type of abnormality cause candidate and the vertical axis as the frequency.

  In the example shown in FIG. 9, the frequency of abnormality D is the highest among the abnormality cause candidates. Therefore, the abnormality cause inference unit 14 identifies the abnormality D having the highest frequency as a true abnormality cause (S5). The abnormality cause (abnormality D here) identified in S5 is output by the data output unit 15 and displayed on the monitor screen by the display unit 16 (S6).

  Next, the creation process (learning) of the conditional probability table H will be described.

  FIG. 10 is a flowchart showing processing for creating a conditional probability table. In the present embodiment, for example, as shown below, the conditional probability table H is created based on the observation data of the alarm corresponding to the cause of the abnormality. That is, first, for example, after setting / specifying each condition node I and each result node O by the operator, the alarm input unit 12 receives the observation data and accumulates the observation data (S11, S12).

  Subsequently, the conditional probability table H (see FIG. 6) is learned and created based on the observation data (S13). Specifically, the value of each condition node I at a certain time is acquired, and the result node O corresponding to the condition node I at that time is acquired, and these acquisitions are performed for a plurality of times. Then, the conditional probability table H is created by deriving the probability of each result node O when the values of the condition nodes I are the same. Finally, the conditional probability table H is output, and the processing for creating the conditional probability table H ends (S14).

  The observation data can be acquired by inputting alarm information from the controller of the CFB boiler and accumulating alarms for each time width. Here, it is preferable to collect as much data as possible, assuming that data accumulated in a certain time width is one data. Incidentally, the timing at which the conditional probability table H is created is, for example, once a year.

  Next, operations and effects of the abnormality cause identification system 10 according to the present embodiment will be described.

  In the abnormality cause identification system 10 of the present embodiment, as described above, an abnormality cause candidate can be obtained from the probability of occurrence of an abnormality cause in a predetermined time width unit, and the abnormality cause can be identified from the frequency of the abnormality cause candidate. . Therefore, even in a CFB boiler in which a plurality of types of alarms are generated, the cause of the abnormality can be specified with high accuracy.

  Further, in the present embodiment, the abnormality cause candidates for each predetermined time width are derived in real time and plotted in the frequency distribution (see FIG. 11), and the frequency change of the abnormality cause candidates with the passage of time can be grasped. . In this case, since the frequency of the true abnormality cause is considered to increase as time passes, it becomes possible to identify the abnormality cause before the actual abnormality occurs, and to take some measures in advance for the abnormality be able to.

  In the present embodiment, as described above, the conditional probability table H of the Bayesian network model is created from the alarm observation data corresponding to the cause of the abnormality, so that it is practical and more accurate and practical. The cause of the abnormality can be identified. Further, as shown in FIG. 6, in the conditional probability table H, the result nodes O corresponding to the values of the respective condition nodes I are represented by probabilities, so that the grounds for causing each result node O to be an abnormality are clarified. . Therefore, by using the Bayesian network model having the conditional probability table H, it is possible to identify a practical abnormality cause that is easy for the operator to use.

  Further, the abnormality cause identifying system 10 of the present embodiment can display not only the cause of the abnormality but also a specific countermeasure by the display unit 16 further displaying information on the countermeasure for the abnormality cause. It becomes possible to converge the abnormal state early.

  Further, as described above, the abnormality cause identification system 10 of the present embodiment can be applied to a CFB boiler as a target for identifying an abnormality cause. In general, in CFB boilers, it is particularly difficult to identify the cause of an abnormality when multiple types of alarms occur. By using a statistical model (Bayesian network model), an appropriate cause can be easily derived. can do.

  The preferred embodiments of the present invention have been described above. However, the present invention is not limited to the above-described embodiments. The present invention is modified without departing from the scope described in the claims or applied to others. It may be.

  For example, in the above embodiment, a Bayesian network model is used as the statistical model, and the abnormality cause inference unit 14 infers the abnormality cause by the conditional probability table H. However, the present invention is not limited to this, and other statistical models (for example, The cause of the abnormality may be inferred using a neural network model, a genetic algorithm model, or the like.

  Moreover, although the said embodiment specified the abnormal cause of the CFB boiler, it is not limited to this, This invention specifies the abnormal cause of other large-scale apparatuses, for example, power generation equipment or water treatment equipment. You can also. Even in this case, the above-described effects are exhibited.

  Moreover, in the said embodiment, although the conditional probability table | surface H was produced based on observation data, it is not limited to this, You may produce by calculation and may produce it by another method. Moreover, in the said embodiment, although the conditional probability table | surface H was created by the abnormality cause identification system 10, you may create the conditional probability table | surface H by another apparatus. Further, when creating the conditional probability table H in the abnormality cause identifying system 10, the abnormality cause identifying system 10 may have a learning unit that learns and creates the conditional probability table H.

  Moreover, in the said embodiment, although 1 second was illustrated as a predetermined | prescribed time width which the alarm | storage accumulation | storage part 13 accumulate | stores an alarm, a predetermined | prescribed time width is not limited to this, In order to exhibit the said effect, it is a candidate for an abnormal cause Any time width that can be obtained appropriately. Moreover, each predetermined time width does not necessarily need to be constant, and may be different time widths.

  DESCRIPTION OF SYMBOLS 10 ... Abnormal cause identification system, 11 ... Model storage part, 12 ... Alarm input part, 13 ... Alarm storage part, 14 ... Abnormal cause reasoning part (inference part), 15 ... Data output part, 16 ... Display part.

Claims (6)

An abnormality cause identification system for identifying an abnormality cause of a device that emits multiple types of alarms,
A model storage unit in which a statistical model related to the occurrence probability of the abnormality cause corresponding to the alarm is stored in advance;
An inference unit for inferring the cause of the abnormality based on the input alarm information and the statistical model;
A display unit for displaying information on the cause of the abnormality inferred by the inference unit,
The reasoning part is
In a certain time, the abnormality cause candidate is obtained based on the alarm information and the statistical model at each of the divided predetermined time widths, and the frequency is calculated for each abnormality cause candidate type,
An abnormality cause identifying system characterized by inferring at least the abnormality cause candidate of the type having the highest frequency as the abnormality cause.   The abnormality cause identifying system according to claim 1, wherein the statistical model is created based on observation data of the alarm corresponding to the cause of the abnormality.   The abnormality cause identification system according to claim 1, wherein the display unit further displays information related to the countermeasure for the abnormality cause.   The abnormality cause identification system according to claim 1, wherein the statistical model is a Bayesian network model. The Bayesian network model has a conditional probability table;
The conditional probability table is
A plurality of condition nodes related to the alarm information, and a plurality of result nodes related to the occurrence of the abnormality cause,
Each value of the plurality of condition nodes is a value related to the presence or absence of the alarm,
At least the value of each of the plurality of condition nodes, the plurality of result nodes, and a random variable are associated with each other;
The reasoning part is
Determine the value of each of the input condition nodes;
5. The cause of abnormality according to claim 4, wherein among the plurality of result nodes associated with the determined value of the condition node, a result node having the largest random variable is derived as the abnormality cause candidate. Specific system. The abnormality cause identification system according to any one of claims 1 to 5, wherein the device is a circulating fluidized bed boiler.

Images