JPS6214210A - Method of analyzing abnormality diagnosis - Google Patents

Abstract

PURPOSE:To make diagnosing of initial indication of abnormal phenomenon by patternizing change of observed signals and making logical processing. CONSTITUTION:In a large-scale system such as a nuclear power plant, when some abnormality occurred, influences of the abnormality are often obserbed in from of change of plural signals from normal state. If change of each observed signal in the direction of increase from normal state is expressed as +1, change in the direction of decrease as -1, and no change is expressed as '0', the state of change of a series of signals A-N can be vector expressed as (a, b, ..., n)=(+1, -1, 0, ..., +1). This state vector analyzes transmission characteristic between signals of object of diagnosis beforehand basing on physical structure of the whole object and experience of analysis of observed data, analyzes the state vector utilizing this.

Description

[Detailed description of the invention]

[Technical Field of the Invention] The present invention relates to an abnormality diagnosis and analysis method for analyzing minute transient changes in a system consisting of multiple signals such as an atomic couplant and extracting information regarding the cause of a phenomenon. (Technical background of the invention and its problems) In a large-scale system such as an atomic couplant,
The changes that occur are transmitted to multiple signals and observed. For this reason, conventionally, in multi-signal systems, a method has often been used to analyze the cause of a phenomenon by integrating the features appearing in a plurality of observed signals. Examples include a method of analyzing the cause of a phenomenon by focusing on the temporal relationship of changes in multiple signals, and a method of estimating the transmission route of an abnormal phenomenon from a pattern of the presence or absence of a change. The above diagnostic method for systems in which multiple process signals can be observed is generally effective when it is clear whether there is a change in the signal, but when the change in item iiA is unclear,
For example, if there is no change in the M measurement even though it is actually changing, or if only a certain signal is observed with a delay, an accurate diagnosis cannot be made. was there. By the way, in large-scale systems such as those mentioned above, early signs of abnormal phenomena often appear as minute changes in some related signals. The following matters are known. That is, (1) it is difficult to distinguish between normal changes (steady fluctuations) and abnormal changes in the signal, and the observed signals are often inaccurate. ■ Changes in the signal are suppressed by the plowing of the feedback loop, so no change appears in the signal that should change. Both of these (1) and ■ lead to uncertainty in observation information, and conventional diagnostic methods are not always effective in diagnosing early signs of transient phenomena or phenomena with minute signal changes. There was a problem that it could not be fixed. [Object of the Invention] The present invention was made in view of the above-mentioned circumstances, and its purpose is to provide an abnormality diagnosis method for a system in which multiple signals can be observed.
The present invention provides an abnormality diagnosis and analysis method that can perform effective diagnosis even for phenomena in which part of the observation information is uncertain, such as the minute change in No. 1. [Summary of the Invention] The present invention provides the following features: - In order to achieve objective 1, extract information regarding the cause of occurrence of a system consisting of multiple signals by analyzing and diagnosing minute changes in multiple signals in a system consisting of multiple signals. In the abnormality diagnosis analysis method, a first means for associating change states of a plurality of signals as a chain of causal relationships between the two signals by combining knowledge regarding an increase/decrease relationship established between the two signals; Even if part of the information in the middle of the chain is lost due to some reason, the hypothesis regarding the change in the signal) is used to compensate for this and estimate the chain of causal relationships. Therefore, the causality between the cause signal of the chain in the customer service level and the result signal of other small units is divided into small units, and the chain of causality is searched for by the second means for each small unit. An anomaly diagnosis analysis method having a third means for analyzing the cause of a phenomenon from the chain of causal relationships between the small m order and the signals within the small order m order obtained as a result of analyzing the relationships. It is related to. Next, the first to third means will be explained. First, to explain the first means, changes in the ffi number can be roughly divided into three states: (increase)/(no change)/(decrease) with respect to each normal state. If these three states are represented by (+1°0, "-1), the signal state will be represented by one of the following nine combinations. ), (+1.-1), (
0,+1), (0,0), (0,-1), (-1,+1
), (-1, 0), (-1, -1) In particular, when a change is transmitted between two signals, -4= There are only the following four combinations. (If there is no change, it cannot be said that the change has been transmitted.) (+1.+
1), (+1.-1), (-1, +1), (-1,-1
) Generally, in a certain system, the transfer characteristics of changes between two signals are determined by the physical structure of the system and can be roughly divided into positive transfer characteristics and negative transfer characteristics as shown below. (Except when there is no transmission.) Positive transfer characteristic (input; increase → output; increase, input; decrease →
Output; Decrease) Negative transfer characteristic (Input; Increase → Output; Decrease, Input; Decrease →
Output; increase) Now, if the transfer characteristics between two signals in a certain system are known, it is possible to determine the causal relationship between the two signals (input → output, cause → effect) by observing the change state of the two signals. Using this, it is possible to describe changes in multiple signals as a chain of causal relationships between two signals. The search for a chain of cause-and-effect relationships basically starts from any observed signal of change and links the chain from cause to effect, or from cause to effect, or r) through trial and error. This is done by expanding the chain and searching for a chain that can explain all the observed changes without contradiction, but if there is known information about causes and effects, it is also possible to use this to narrow the scope of the search. It is. By representing the states of multiple observed signals as a chain of causal relationships in this way, it becomes possible to divide the observed phenomenon into groups of change-causing signals and resultant signals, which is useful for analyzing the causes of phenomena that occur. γ one. In a real system, the number of observed signals is large and the relationships between signals are complex, so the chain of causal relationships that can explain the states of all the signals w4il111 is limited, but there is no single chain of causal relationships. Even if there is no such information, it is possible to narrow down the possible causes, which is useful for analyzing the cause of the phenomenon that has occurred. Next, to explain the second method, the first method, which analyzes the observed phenomenon described above as a chain of causal relationships (4), assumes that the state of change of each signal can be observed. Therefore, as with conventional methods, if the observation of the signal is uncertain, it is not possible to make an effective diagnosis. Therefore, in the process of searching for a chain of causal relationships, we (For issue 4, it is already l17f)
Hypothesizing a change that does not contradict the chain established based on the a+++ data, continue searching for the chain using ~γ.
In the subsequent processing process, if there is no contradiction with observed data, the hypothesis is treated as correct. By using such a method, some IrV i+
Even if no change is observed, the chain of causality can be searched for without breaking, and an effective diagnosis can be made. Finally, to explain the third means, the second means
Introducing hypotheses as in the method of Therefore, by discarding such unreliable search results midway through, meaningless searches can be avoided and expansion of the search space can be suppressed. However, when the chain is long, the rate at which hypotheses are introduced is higher than when it is short, so if there are chains of extremely different lengths in the search space, it becomes difficult to uniformly evaluate reliability. Therefore, in this method, the length of the chain searched in a series of processes is limited by dividing the diagnostic target into several meridians according to its physical structure and searching for the chain for each unit. Facilitates uniform reliability evaluation for chains in which hypotheses are introduced. Furthermore, in this method, after searching for a chain of causal relationships for each small unit, a chain is established between the signal estimated to be the cause of change at each small change position and the resulting signal at another small threat position. analyze the causal relationships between small units. By analyzing the causal relationship in two stages in this way, it is easier to evaluate the reliability of the search results, and at the same time it is also possible to analyze cases where some kind of abnormality occurs at the same time in two different small ridge positions. becomes. Furthermore, by dividing the search for a chain of causal relationships into a number of unique units, it is possible to improve processing efficiency by parallel processing. [Embodiment of the Invention] An embodiment of the present invention will be described with reference to the drawings. In large-scale systems such as atomic couplants,
When an abnormality occurs, its effect is often observed in the form of multiple signals changing from their normal state. For each *1+1 signal, a change in the increasing direction from the normal state is set as +1, a change in the decreasing direction is set as -1, and when no change is observed, it is set as 0], then the series of signals (A-N) is The state of change is (a+ b+ Cr "'r n)" (+1+
L The standard value in this case can be the set value or target value of the control system, or the predicted value by an autoregressive model, adaptive filter, or physical model. It is possible to use various values.Also, U
In addition to directly creating a state vector from changes in ν series data, it is also possible to create a state vector from data such as differential signals and spectra, and the present invention can be applied to both of them. . In the first means of the present invention, the transfer characteristics between signals of the diagnostic target are analyzed in advance based on the physical structure of the entire target and experience in analyzing observation data, and the state vector is analyzed using this. Now, as a simple example, consider a system in which 64.1';+ of (A -1. (Relation;]) A-) B (Positive transfer characteristic) (Relation; 2)
F(-+A (transfer characteristic of iF) (relationship; 3) B→0
(Positive transfer characteristic) (Relation; 4) C→I) (Positive transfer characteristic) (Relation; 5) 1
)→■・: (Positive transfer characteristic) (Relationship; 6) D→I” (
Positive transfer characteristic) (Relation; 7) E → [] (Positive transfer characteristic)
(Relation; 8) F → 11 (Negative transfer characteristic) In this system, the change in Araka is 5ill 1llll+ and the state vector (a + b + c + 40 + f) - (+ I + + l+, +1, +1.0) - I state ;1) is obtained,
a = + ] as the starting point

[2] When analysis starts from the cause to the effect, b-+1 is estimated by applying (Relation: 1). Go (ha, ail) Since it does not contradict the i1M fact, we preserve the relationship and then perform the analysis starting from b = +1. At this time, (relation; 2) and (relation;
3) is applicable, and applying either relationship will not contradict the observed data, but if (relation: 2) is applied, an infinite chain will occur (A → 13 → A →...). Since there is saturation, we reject this and record A-+B-+C) as the correct chain. If there are any applicable branches, each is searched sequentially. By repeating this search process, the previous state vector can be interpreted as an increase in signal A and a causal chain of (A→t(→C→D→1・;)). In this example, the search is performed from the cause to the effect, but conversely, the chain of causal relationships from the effect to the cause (T:→D ->
It is also possible to search for C-)B-+A). When starting the search from an arbitrary signal, 5, for example (0→
BAA), first search in the direction of cause lj to select signals vf (A and mouth in this case) that can be the cause, and then search from those (, 1 to (A→H→C) →D→E),
(R→C→1)→E), search is performed in the direction of the result, and the series #(A→■1→C−+l) that satisfies all observation data is performed.
) → 1 b), or a method of searching in the j cause direction from all the signals for which changes are observed, and selecting the signal that is the cause common to all the signals as the cause. etc. are possible. However, with the above means, if part of the observation information is missing, for example, changes in signals C and D are not observed in (state; ]) for some reason, and the state vector is changed to the following (state; 2) or Suppose that the state is (state; 3). (alblcrd161f) = (+], +I, O, +1
．． +1IO)・(state;2)(a+b+c+d, s, f
) = (R, Published, 0. O, Old, 0) (Status; 3) In this case, all chain searches end at (A→13) and no correct result is obtained. Therefore, in the second means of the present invention, C=+1 is estimated from b=+1 by applying (relationship; 3). On the other hand, when it is confirmed that the observed fact is Q=0, the search is continued by introducing the estimated result C=+1 as a hypothesis without reducing the search. (See Figure 2) For example, (state 53
), d=+1 is estimated for hypothesis C=+1, but since the observed fact is d=0, hypothesis d=+1 is introduced again. Here, if d=-1, hypothesis d=+1 is not introduced and the chain ends at (A→B −+ (7,). If no contradiction between the hypothesis and observed data is found, these hypotheses are By considering it as true, (A→B-+C-+r') in both (state: 2) and (state: 3)
→E) chain is obtained. In this way, even if some observational information is missing, by filling in the missing information with a hypothesis, it is possible to continue searching for a chain of causal relationships. However, if we introduce such a hypothesis unconditionally,
For example, the change in only the signal is 11811! In this case, all possible combinations are searched and the chain search space expands because none of the hypotheses is inconsistent with the l1llll data. Therefore, in the present invention, as shown below, in one search process, when the number of hypotheses is large compared to the number of observed facts as in (:3), or when hypotheses are continuous as in 0), etc. It is assumed that the reliability of the estimation results is low, and by discarding such low reliability results during the search, meaningless searches are avoided and expansion of the search space is suppressed. (3) A→-1-+dan→D-+9→ENG→G (lower case; hypothesis) (a) A→B→-9-→,! l! −→9−→F→G
(Koshiji; Hypothesis) However, for example, (A→B −* CH
2) When the chain is long (→r<→F→G→H→...), the weight of one hypothesis relative to the overall reliability is greater than when the chain is short (A-+B→C). Therefore, if there are chains of extremely different lengths and shorts in the search space, it becomes difficult to uniformly evaluate reliability. For this reason, in the third means of the present invention, the diagnostic target is divided into several small units according to its physical structure, and a chain search is performed for each small unit.1Thus, a series of processing By limiting the length of the chain of causal relationships explored in After IIQing the chain of causality for each small unit, for the signal estimated to be the cause of displacement in each small unit, by checking whether a chain with the result No. 48 in other small units is established) γ. , Analyze the causal relationship between small threat positions. For example, as shown in FIG. 3, the diagnostic targets are x and y. It is divided into three small units of 2, and the causal relationships within each are determined using the method described above.
yo→y1→...→VII)+(Zo→z1→...
→Zn), if it is estimated that →ν. +VJ
""Zo, (i = t ~ Ω), nj (j = l ~
If m) exists, then X→Y−+Z as in (eS:1)
A causal relationship between the small fist positions is established, and the observed change in the signal is the result. It can be assumed that this was caused by a change in Also, if work 1 (i = 1
~Q) + yj (, i=I~m) exists, then (ax,
As shown in 2), the causal relationship between the positions of X→Z, Y-+Z is established, and the observed change in the signal is
y. It can be considered that this may be caused by changes in the two signals. As described above, the present invention is characterized in that data can be analyzed even when abnormalities occur simultaneously in a plurality of different small units. Next, one embodiment of the present invention will be described using the flowchart of FIG. First, the first signal sequence S1 will be explained. In response to a search start command, a signal from the plant is input as a first step 1. In the second step 2, if it is determined that there is no search start signal, that is, the search has ended, it is input to the selection of the causal element in the tenth step 10.6 If it is determined that there is a search start signal, it is input to the next third step 3. It is determined whether an applicable relationship is selected. If it is determined that there is no applicable signal, the process returns to the second step 2, and if it is determined that there is an applicable signal, it is input to the fourth step 4, where the propagating signal change is estimated and then the next step is performed. 5 is input to step 5. In this fifth step 5, a comparison is made with the observed signal, and if it is determined that there is no change in the observed signal, the process is performed manually in a sixth step 6, where a hypothesis is set. Also, if it is determined that there is an observed signal change, the next step 7 is performed.
and is compared with the observed signal. and,
If it is determined that the estimation result and the observed data are inconsistent, the process is inputted to an eighth step 8 and the above-mentioned hypothesis is discarded. If it is determined that the estimation result and the am data do not contradict each other, the relationship input and applied to the ninth step 9 is recorded. Therefore, the outputs of the sixth step 6, seventh step 7 and ninth step 9 are input to the third step 3 and are repeated until it is determined that there is no applicable relationship. If it is input to step 2 and it is determined that there is no search start signal, it means that the search has ended, and the search is input to the tenth step 10 as described above. In the 10th step 10, the causal element is selected and the next 11th step 1
1a, 11b, and llc. This eleventh step 11a. In steps 11b and llc, the inclusion relationship between cause elements and effect elements between small units is searched, and the results are used in the next 12th step 1.
2, and in the twelfth step 12, the phenomenon due to the causal relationship of signal changes is characterized. Next, the second signal sequence S2 is also subjected to signal processing similar to that of the first signal sequence S1. Sunao) also this second one?
In the f1 series, each step corresponding to the first signal series S1 is marked with a dash, but since the processing content is similar to that of the first signal series S], a detailed explanation thereof will be omitted. do. In this second signal sequence S2 as well, the cause element is selected in the tenth step 10', and the inclusion relationship between the cause element and the effect element between small units is searched in the next eleventh step lla, llb, llc. After that, in the next twelfth step 12, phenomena caused by the causal relationship of signal changes are characterized. [Effects of the Invention] As explained above, according to the present invention, by patterning the changes in the wt-measured signal and performing logical processing,
Even when individual signals do not have clear characteristics, it is possible to obtain information about the cause of the phenomenon from the minute characteristics that appear in multiple signals, making it possible to diagnose early signs of abnormal phenomena. Furthermore, by dividing one diagnostic target into several independent degrees and analyzing them, it becomes possible to diagnose abnormalities due to multiple causes.

[Brief explanation of drawings]

FIG. 1 is a flowchart for explaining an abnormality diagnosis analysis method according to an embodiment of the present invention, FIG. 2 is a diagram for explaining a state vector according to the present invention, and FIG. 3 is a diagnostic target according to the present invention. It is a diagram for dividing into small units and explaining the causal relationship for each small unit. 1. Signal human power 2. Selection of search start signal 3. Estimation of applicable relationships 4. Estimation of propagating signal changes 5.7. Verification with observed signals 6. Hypothesis Setting 8
...Discarding the hypothesis 9...Recording the applied relationship 10...Selecting the causal element 1.1...Searching for the inclusion relationship between the cause and effect elements between platoon positions 12...Causal relationship of signal change For the characterization of the phenomenon, please contact my agent, patent attorney Kodo Inomata, and Hiroko Inomata. Q 10 ÷ + + ÷ He ゝq+ (1 prisoner) (5 prisoners) (421) χO→(χi, i near, 1)↓ ′jO-(:/, 7゜甚X, Yyuヱ(e7- , ln 1o −(χi, t=/
]↓
"(Got, j=/-m) → lo → (inkstone = 7, suj' = 1 ~ illusion ↓ 'lo" (14 @, f, = Izn-) n 7hiko = lυri 3rd Figure-raR-

Claims (1)

[Claims] In an abnormality diagnosis and analysis method that extracts information regarding the cause of occurrence of a system by analyzing and diagnosing minute changes in multiple signals in a system consisting of multiple signals, it combines knowledge about the increase/decrease relationship established between two signals. a first means for associating change states of a plurality of signals as a chain of causal relationships between two signals; A second means for supplementing this by forming a hypothesis and estimating a chain of causal relationships; and after dividing the diagnostic target into small units according to their functions and searching for a chain of causal relationships for each small unit by the second means. Analyze the causal relationships between the chain of cause signals in each small unit and the result signals of other small units, and use the resulting chain of causal relationships between the small units and signals within the small unit to determine the phenomena. An abnormality diagnosis analysis method characterized by having a third means for analyzing the cause.