JPH04113220A - Method and device for identifying abnormal event - Google Patents

Abstract

PURPOSE:To improve identification accuracy by identifying an abnormality using data before and after the occurrence of abnormality. CONSTITUTION:A data device for identifying even 110 selects data which is needed for identifying an event from a time-series data in reference to an abnormality signal generation time with a previously determined specific time width according to a type of an abnormality signal 30. A neural network 130A takes in neural network data which has already been learned from an external memory device 140 and establishes the neural network, and obtains a certain output pattern by the neural network processing using data for identifying the event. A processing part 130B outputs an identified result 160 by referring to an output pattern/data table for abnormality event which are memorized within an external memory device 150.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to an abnormal event identification method and apparatus for identifying abnormalities in abnormal event identification targets such as plants and machines.

[Conventional technology]

In order to maintain the health of a plant or machine, it is necessary to take early countermeasures against abnormal events caused by failures or malfunctions of various equipment, operator errors, external disturbances, etc. For this purpose, abnormal events must be detected and identified at an early stage. For example, pattern recognition technology is used to identify abnormal events from various state variables of a plant or machine.

Recently, the effectiveness of neural networks that simulate biological neural networks for pattern recognition has been confirmed. This neural network has the excellent characteristics of high-speed processing and the ability to recognize patterns even if some noise is mixed into the input signal.

On the application of neural networks to the identification of transient anomalous events in atomic couplants, Nis°P.

IE, Volume 1095, Applications of Artificial Intelligence 7 (1989) Paragraphs 851 to 856 (SPIE, V
ol, 1095°Applications of A
rtificial Intelligence (1
989) discussed in PP851-856).

This paper states that since the output pattern of each detector in a plant is uniquely determined for each abnormal event, it can be used as information to identify the state of the plant at any given time.

The authors also state that the effectiveness of neural networks in the application of abnormality diagnosis has been confirmed in terms of noise resistance and real-time processing.

[Problem to be solved by the invention]

In devices that constantly capture time-series data from detectors attached to actual plants or machines to identify abnormal events, the time of abnormality occurrence is not clear, so the time series input to the neural network is It is conceivable that the data patterns will differ considerably, and the accuracy of abnormal event identification by the neural network will deteriorate.

Therefore, it is necessary to use some method to determine the time at which the transient abnormality begins, and to generate data based on that time.

Furthermore, due to differences in transient abnormal events, it is conceivable that the temporal behavior unique to the event may change.

In that case, it may be difficult to identify all events using time series data captured at equal time intervals.

An object of the present invention is to provide an abnormal event identification method and apparatus for identifying an abnormal event according to the time of occurrence of the abnormality (claims 1 to 1.1).

A further object of the present invention is to provide an abnormal event identification method and device that identifies an abnormal event by reflecting the content of the abnormal event (claims 2 and 5).

A further object of the present invention is to provide an abnormal event identification method and apparatus that identify abnormal events using a neural network (claims 6 to 11).

[Means to solve the problem]

The identification method of the present invention uses signals from a system to be identified for an abnormal event before and after the abnormal signal is generated as data for abnormality identification (claim 1).

The identification method of the present invention selects signals from a system to be identified for abnormal events before and after the abnormal signal generation time at a sampling time corresponding to the type of abnormal signal, and uses the selected signals as data for abnormality identification. (Claim 2).

The identification device of the present invention includes a recording means for recording a detection signal from a system to be identified as abnormal event as time-series data, a detecting means for detecting an abnormal signal occurrence time in the abnormal event identification target system, and a reference to the abnormal signal occurrence time. means for reading out time-series data in the recording means existing in a predetermined time width before and after the time as event identification data; and identification means for identifying an abnormal event from the event identification data. (Claim 3)

The identification device of the present invention includes a recording means for recording a detection signal from an abnormal event identification target as time series data and a logical signal from the abnormal event identification target as logical data, and an abnormal signal occurrence time of the abnormal event identification target system. a detection means for detecting;
means for reading out time-series data and logical data in the recording means that exist in a predetermined time width before and after the abnormal signal generation time as event identification data; and from the event identification data. an identification means for identifying an abnormal event;
(Claim 4)

Furthermore, in the identification device of the present invention, the reading means includes means for specifying a predetermined time width before and after the abnormal signal generation time, and a sampling time corresponding to the type of abnormal signal is stored in advance. storage means for detecting the type of the generated abnormal signal and reading the corresponding sampling time from the storage means; and means for reading out identification data (claim 5).

Further, in the identification device of the present invention, the identification means includes a neural network (claim 6).

Further, in the identification device of the present invention, the neural network has a pre-learning function based on a simulator (claim 7).

Furthermore, in the identification device of the present invention, during the learning, the sampling time width and the number of data points of data to be provided to the input layer, which is the input part of the neural network, are changed in accordance with the abnormal signal (claim 8).

Further, in the identification device of the present invention, the number of layers of the neural network and the number of units that are units for processing information in each layer are given from an external memory (claim 9).

Furthermore, in the identification device of the present invention, the type of data used in the neural network,
It has an inquiry function that can display the number of data points, sampling time width, and trained neural network data (claim 10).

Furthermore, the identification device of the present invention has a function of performing knowledge processing using a database in which knowledge corresponding to abnormal events in plants or machines is registered, and determining more detailed abnormal items (claim 11).

[For production]

According to the present invention, the signals from the abnormal event identification target system before and after the abnormality occurrence time are used as data for abnormality identification, thereby improving the identification accuracy (claims 1 and 3.4).

According to the present invention, data before and after the occurrence of the abnormality are selected at a sampling time corresponding to the type and rate of change of the abnormal signal, and abnormality identification that matches the type and rate of change is achieved (claims 2 and 5).

According to the present invention, abnormality identification is achieved using a neural network (claims 6 to 11).

〔Example〕

FIG. 1 shows an embodiment of the abnormality identification device of the present invention.

The components shown in FIG. 1 consist of the following.

Plant 10: This is a plant that is a target for abnormal event identification.

Detection signal and logic signal 20: A detection signal from a detector (not shown) installed in the plant 10 and a logic signal indicating the operating state of each device in the plant.

Abnormal signal 30: An abnormal signal generated when the plant 10 is abnormal.

Data capture/recording device 40: Captures and records the detection signal of the detector and the logic signal 20.

Simulator 60: Simulates abnormal events and time-series data and logic signals of the detector at abnormal times.

Data 90: data of simulated abnormal signal generation time, time series data before and after the simulated abnormal signal generation time, and data of logical signals.

Data 100: data of abnormal signal generation time in an actual plant, time series data before and after the abnormal signal generation time, and data of logical signals.

Based on the event identification data device 110 and the time of occurrence of the abnormal signal 20, data of a predetermined time width predetermined according to the type of the abnormal signal 20 is selected to generate data 120 necessary for event identification. It is something to do.

In addition to the coral network 130A, the event identification section has a processing section 130B that performs learning processing and identification processing) 130
After the learned neural network data is imported from the external storage bag W (memory) 140 to establish a neural network, the event identification data 120 is imported to identify an abnormal event. Furthermore,
In order to obtain the trained neural network data, the simulated data 90 generated from the plant dynamic characteristic model simulator 60 is imported, the neural network is trained, and the results are stored in the external storage device 14.
It has the function of storing 0. Identification of the abnormal event is performed using 130 output patterns of a neural network using event identification data as input, and an external storage device (memory) 1.
This is done by referring to a table of output pattern/abnormal event correspondence data stored in 50. signal 160
shows the identification results.

Display terminal device 170: A CRT display and terminal device for displaying the identification results 160. It has a terminal section such as a keyboard.

Signals 180, 190, 200 are signals for selecting whether the abnormal event identification of this embodiment is to function in the learning mode or in the identification mode. That is, when the learning mode is specified, the simulator 60 is operated, the resulting simulated data 90 is sent to the event identification data device 110, where the simulated event identification data 120 is obtained, and the neural network 130A performs learning. The signal 1 is used to perform control to store the results in the external memory IJ14.
80.190.200 indicates.

On the other hand, when the identification mode is specified, various learned data obtained in the above learning are read from the external memory 140 and set in the neural network 130A, and various data 100 from the data acquisition/recording device 40 are used as event identification data. The event identification data 120 obtained here is sent to the neural network 130A, and the neural network 130A performs event identification using the table in the external memory 150 using the signals 180 and 190.゜200 indicates.

Here, the signal 180.190.200 is
This is given by the user's (operator's) instructions.

Next, FIG. 2 shows a processing flowchart of the embodiment of FIG. 1. First, it is determined whether the user's request is a learning mode or an identification mode (processing if). If it is the learning mode, processes 1g-1 to 11-3 are executed.

If it is the identification mode, processes 18 to 1e are executed.

The learning mode or the identification mode is determined by the user's request at the time, and may be specified by the user on the character screen of the CRT 170 (or specified using the keyboard).

First, in the case of learning mode (processing 1g-1 to l 1-3)
Explain. Start the simulator 60 and start the plant 10
The dynamic characteristic model is operated to generate simulated time-series data of various detector signals for a certain abnormal event and data 90 of simulated logic signals at that time (processing Ig-1).

This data 90 is sent to the event identification data group W110, and event identification data is generated (1c). This generation process is the same as the process in identification mode (described later).

In the neural network 130A, event identification data 120 based on data 90 is given as input data, output patterns corresponding to abnormal events are given as training data from an external memory 150 storing an output pattern/abnormal event correspondence data table, and normal The neural network 130A is made to learn using the error backpropagation algorithm (processing 1i-1).

Process to change the next event to be simulated to another event 1 g-2
), a series of processing (processing 1 g-1, 1c, 1 i-1
゜1 g-2) is repeated as many times as the number of events to be identified. Furthermore,
In the neural network 130A, the above-mentioned processes are repeated until predetermined learning termination conditions (square error small, completion of a predetermined number of learning times) are met.

When the learning end conditions are met, the learning results are stored in external memory 1.
40, the process 11-3) returns to the initial state, the step of listening to the user's request (processing If).

Next, the case of the identified mote will be explained according to the processing (processing 18 to le).

First, while the abnormal signal 30 is not generated and the abnormal signal 30
Even if this occurs, until a predetermined period of time has elapsed (Process 1a)
, the signals of the detectors installed in the plant 10 and the logic signals 20 representing the operating status of the plant are taken in as data.
The data is captured and recorded in the recording device 40 (processing 1b).

A predetermined period of time has passed since the abnormality signal 30 was generated (processing 1
a), time series data before and after the abnormal signal generation time, abnormal signal generation time data, and logic signal data 100 are sent to an event identification data device 110.

The event identification data device 110 selects data necessary for identifying an event from time-series data in a predetermined time width according to the type of the abnormal signal 30, using the abnormal signal occurrence time as a reference.

The data generated from this time-series data and the logical signal data are put together as event identification data 120, and the event identification data group W110 is sent to the neural network 130A (processing 1c).

In the neural network 130A, already learned neural network data such as weighting coefficients and threshold values obtained as a result of learning are stored in the external storage device 1.
40 to establish a neural network,
Using the data for event identification, this established neural network performs neural network processing,
A certain Ryo Kabata turn is obtained.

Regarding this output pattern, the processing unit 130B outputs the identification result 160 by referring to the output pattern/abnormal event correspondence data table stored in the external storage device 150 (processing 1d). And identification result 16
0 is displayed on the display terminal device 170 (processing 1e).

In the case of the learning mode, if an abnormal event occurs in the plant 10, this learning mode is automatically ended and the mode shifts to the identification mode.

FIG. 3 is a diagram showing an embodiment of the data capturing/recording device 40. As shown in FIG.

2OA: A signal from a detector installed in the plant.

20B: A logical signal indicating the operating state of the plant equipment. In the example of a nuclear power plant, a logical signal is (a) a logical signal indicating a water pump trip (tt 1 sr if trip is O''') (b)
Logic signal indicating high water level (It l F+ if the upper limit water level threshold is exceeded, l (0) if it does not exceed it) (c) Logic signal indicating high vapor pressure (“1” if the upper limit vapor pressure threshold is exceeded;
If not exceeded, it is "O") (iv) A logic signal indicating that the bypass valve is open (It I It when it is open, u OPI when it is closed).

In Figure 3, 30: Abnormal signal generated when a plant abnormality occurs.

100: data of the abnormal signal generation time in the actual plant, time series data before and after the abnormal signal generation time, and data of logical signals.

40a: A recording medium for recording various data from the plant.

40a-1... Time series data of various detector signals of the plant D+DI2+...t D21+ D22+...
...A time-series data recording unit that records endlessly.

40a-2: An abnormal signal generation time recording unit that records the abnormal signal generation time.

40a-3...A logic signal recording unit that records logic signals LI+L2*... generated from plant peripheral devices and indicating the operating state of the plant.

40b... A clock for referring to time.

40c: An abnormal signal generation time calculation device that calculates the time t1 of the abnormal signal 30 generation by referring to the time in the clock 40b.

40d. A time series that determines the recording time of the detector signal after the occurrence of the abnormal signal 30 by comparing the time t of the clock 40b and the abnormal signal generation time recorded in the abnormal signal generation time recording section 40a-2. It is a data recording limited device.

Next, the operation of the apparatus shown in FIG. 3 will be explained using the flowchart shown in FIG.

First, until the abnormal signal 30 is input to the abnormal signal generation time calculation device 40c in the data acquisition/recording device 40, (
Process 4a), time series data recording unit 40 of recording medium 40a
The detector signal 20A and the logic signal 20B are recorded in the a-1 and logic signal recording sections 40a-3, respectively, together with the time of the clock 40b (processing 4b). The logic signal recording section 40a-3 is a memory prepared for only the types of logic signals.

When the abnormal signal 30 is input to the abnormal signal generation time calculation device 40c (processing 4a), the time when the abnormal signal 30 is input is determined with reference to the clock 40b, and the time is recorded in the abnormal signal generation time recording section 40a-2. (Process 4c).

A time T for recording the detector signal from the occurrence of the abnormal signal 30 is determined in advance, and the time-series data recording limiting device 40
Record it in advance in d.

The elapsed time from the abnormal signal generation time recorded in the abnormal signal generation time recording section 40a-2 is calculated from the time indicated by the clock 40b, and as long as the elapsed time does not exceed the pre-recorded value T, the detector signal 20A is and the logic signal 20B, and the recording medium 40a along with the time of the clock 40b.
(Process 4d).

4b).

When time T has elapsed from the abnormal signal generation time (processing 4d), recording of the signals 20A and 20B is stopped (processing 4e).
.

Thereafter, data 100, which is a collection of time-series data of detector signals before and after the recorded abnormal signal generation time, data on the generation time of the abnormal signal 30, and logical signal data, is sent to the event identification data device 110 (processed 4f).

FIG. 5 is an example diagram of the event identification data device 110.

110a: A time-series data sampling device that selects data to be provided to the event identification device from time-series data.

110b: An external storage device that records sampling times corresponding to the abnormality signal 30 or simulated abnormality signal 31 when extracting data necessary for event identification from time-series data. The sampling time is Δj++
A large number of values are set, such as Δj2+..., Δ1o, and the sampling time corresponds to the type of abnormal signal.

110c: A learning/identification data changeover switch for selecting whether the data 90 from the simulator is the data 100 from the plant.

110a-1: A memory for storing data necessary for identifying events generated from time-series data of detector signals.

110a-2: Logic signal recording section.

110cm1...Logic signal recording section.

110cm2: Time series data recording section.

110cm3: Abnormal signal generation time recording section.

Next, using Figure 5 and Figure 6, the event identification data device 1 in Figure 5 is
The operation of 1110 will be explained.

First, if the request from the user through the display terminal device 170 is for learning mode (processing 11a), the learning/identification data changeover switch 110c is switched to learning mode via the signal line 190. The data input to the event identification data device 110 becomes the simulated data 90 from the simulator 60 (processing 11c).

If the request from the user is not learning mode (process 11a
), the switch 110c is switched to identification, and the input data becomes the actual plant data 100 (processing 11b).

Input data is simulated data 90, actual plant data 100
In any case, the respective time series data is loaded into the time series data recording section 110cm2, the logic signal data is loaded into the logic signal recording section 110cm1, and the abnormal signal generation time data is loaded into the abnormal signal generation time recording section 110cm3 (processing 11d). .

The time series data sampling device 110a selects one of the sampling times stored in the external storage device], ], Ob depending on the type of the simulated abnormal signal 31 or the abnormal signal 30 (processing 11e).

Data is extracted from the time series data at the selected sampling time based on the abnormal signal generation time and stored in the memory 110a-1 (processing 11f).

The extracted data and logic signal data 120 are sent to the event identification device 130 (processing 11g).

The data sampling time was changed depending on the type of abnormal signal, and a specific example of this will be described below.

(b) There is a way to do this depending on the source of the abnormal signal.

In this case, it is assumed that the location where the abnormal signal occurs can be identified. However, knowing the source does not necessarily lead to or contradict the identification of the abnormal event. Rather, in a large system such as the yK child couplant, abnormalities appear as various phenomena, and it is absolutely necessary to identify the abnormalities from these phenomena.

(b) There is a way to observe changes in the detector signal. The characteristics of the detector signal can be defined by its amplitude, period (a kind of rate of change), etc. Anomalies can be identified based on the magnitude of these characteristics (such as whether the amplitude is small or large, or whether the rate of change is fast or slow).

(c) A combination of (a) and (b) can also be used to identify abnormalities. That is, the sampling time is selected depending on the source of the abnormality and the characteristics of the abnormal signal.

FIG. 7 is a diagram illustrating the operation of the identification data device when the sampling time is changed in accordance with the rate of change of the detector signal or when the type of abnormal signal is changed.

When abnormal signal A occurs as shown in Fig. 7(a),
Assume that there is a transient response that causes a fast change such as 40a-10.

At this time, judging that it is such a fast change, Δt1
Select a short sampling time like
+... is collected.

When abnormal signal B occurs as shown in Fig. 7(b),
Assume that there is a transient response that causes a slow change such as 40a-11.

At this time, judging that it is such a delayed change, Δt2
A long sampling time is selected as shown in FIG. 2, and identification data b')2+b3 is collected at this sampling time.

By introducing the sampling time according to the abnormal signal as described above, it is possible to capture transient abnormal events that change uniquely over time, and it is possible to reduce the rate of misrecognition when identifying events.

FIG. 8 shows an example diagram of the event identification device 130.

130A: Neural network.

130B1.130B2... constitutes a processing section 130B, 130B1 is an output pattern/event comparison section, 130B
2 is an identification/learning switching section.

140a: trained neural network data such as the number of neural network layers, the number of units in each layer, weighting coefficients for connections between units, and threshold values for each unit.

140b: Neural network data immediately after learning.

140c: input section of the neural network.

140d: Output section of the neural network.

Next, the operation of the event identification section 130 shown in FIG. 8 will be explained using the flowchart shown in FIG. 9.

First, it is determined whether the user's request 180 is a learning mode or an identification mode (processing 13a). When the user's request through the display terminal device 170 is the learning mode, the identification/learning mode switching unit 130B2 is set to only flow data on the learning side through the signal line 180.

The output turn 130a corresponding to the abnormal event simulated by the simulator 60 is converted into a cover turn/abnormal event response data.
The table is provided from the output pattern/abnormal event comparison unit 130Bl to the output unit 140d of the neural network 130A (processing 131).

The event identification data 120 generated from the simulated data 90 from the simulator 60 is taken in and applied to the input section 140c of the neural network 130A (processing 13j).
.

After this, the neural network 130A is trained and the weighting coefficients and threshold values are corrected (Process 13k), and if the predetermined learning end conditions are not met (Process 13Q).
), another abnormal event is selected and processed 13m), the simulator 60 is activated, and new event identification data 120 is created from the simulated plant data.

After that, in processing 13i, 13j, 13, 1.3M
, 13m is repeated until the learning termination condition is met. When the learning end conditions are met (process 13
The neural network data 140b after learning, such as pieces, weights, and thresholds, is stored in the external storage device 140 (process 13n). The operation then returns to the starting point of execution of the event identification device.

Next, when the user's request through the display terminal device 170 is not the learning mode (processing 13a), the identification/learning mode switching unit 130B2 is set to only flow data on the identification side via the signal line 180. First, the learned neural network data 140a is transferred to the neural network 1.
30A (processing 13b). After this, abnormal signal 3
3c) When the event identification data 120 is generated, the event identification data 120 is taken in (processing 13d).

Next, the output pattern 130a is obtained by the neural network processing in the neural network 130C (processing 13e).

Data is fetched from the output pattern/abnormal event correspondence data table 150 into the output pattern/abnormal event comparison section 130B1 (process 13f), and the identification result 160 is obtained by the output pattern/abnormal event comparison section 130B1 (process 13g).

The identification result 160 is then displayed on the display terminal device 170.

According to this embodiment, trained neural network data can be obtained during learning, and this data can be used during identification to identify abnormal events.

In addition, in the abnormal event identification device in this embodiment, when operating in the learning mode, the terminal device displays the sampling time width and data points of data to be applied to the input layer, which is the input part of the neural network that constitutes the event identification device 130. It can also be changed via 170.

In addition, the display terminal device 170 can display the type of data used in the abnormal event identification device, the number of data points, the sampling time width, and the trained neural network data in response to an inquiry from a user.

An example of the output turn/abnormal event correspondence data table 150 is shown in FIG. Three examples of abnormal events were shown for nuclear couplers: main steam isolation valve closure, loss of feedwater heating, and loss of total feedwater flow. - On the other hand, the output pattern shows an example of binary data. That is, the output from the neural network is often binary data of 1 and 0, and the figure shows an example of a 3-bit output.

Note that during learning, output patterns corresponding to events are given as training data to the neural network.

According to the present embodiment described in FIGS. 1 to 10 above, the following effects are achieved.

(a) According to this embodiment, since the event identification data is created based on the abnormal signal generation time, the identification data corresponding to each abnormal event has 22 different patterns. As a result, misrecognition of events is reduced, and even complex patterns of identification data can be identified.

Furthermore, by changing the data sampling time depending on the type of abnormal signal, it is possible to capture temporal changes specific to abnormal events, and it is possible to reduce erroneous recognition of events.

(b) By having a simulator that can simulate abnormal transient changes in a plant, the neural network within the event identification device can be trained. Therefore, even if there is a change in specifications on the plant side, the event identification device can respond quickly.

(C) According to the abnormal event identification device of this embodiment, in addition to the time series data from the plant, data regarding the occurrence or non-occurrence of a logical signal indicating the operating state of the plant 2 is used. Since the information regarding the abnormality becomes clearer, abnormal events in the plant can be identified with high accuracy.

(d) Sampling time width of data given to the input layer of the neural network during learning using a simulator;
The number of data points can be changed.

Therefore, it is possible to provide the event identification device with data extracted at the optimal sampling time from time-series data of unusual events that change over time, and it is possible to reduce the rate of erroneously recognizing events.

(E) Regarding the neural network in the event identification unit 130, data on the number of layers thereof, the number of units that are units for processing information in each layer, and the corresponding trained neural network data can be provided from the external storage device M140. can.

Therefore, the configuration of the neural network used in the event identification device 130 can be easily updated, and maintenance of the device becomes easy.

(f) In response to a user's inquiry, the event identification unit 130
Since the type of data used, the number of data points, the sampling time width, and the trained neural network data can be displayed on the display terminal device 170, it is possible to confirm the data used, and the reliability of the abnormal event identification device is increased. It increases.

FIG. 11 is a block diagram showing the configuration of an abnormal event identification device that is a modification of the first embodiment.

The changes from the first embodiment include a new data flow 10 from the data capture/recording device 40 to the simulator 60.
1 is added, and the knowledge processing device 2], 0, and the abnormal partial knowledge base 220 are added.

The data 101 is the same as the data 100 of the abnormal signal occurrence time in the actual plant, the time series data before and after the abnormal signal occurrence time, and the logical signal data 100.

The abnormal part knowledge base 220 is an abnormal part knowledge base that stores knowledge about detailed abnormal parts corresponding to various abnormal events.

The knowledge processing device W210 is a knowledge processing device that uses the identification result 160 from the event identification unit 130 and the abnormal portion knowledge base 220 to determine a more detailed abnormal portion of the plant.

Next, the operation of the abnormal event identification device of this embodiment will be explained based on the flowchart of FIG. 12.

In FIG. 12, the operation in the case of the learning mode (processing if) is exactly the same except that the actual machine data is also output in processing 1g-10, so the explanation will be omitted here.

On the other hand, if the user's request is not in the learning mode (processing if), the abnormal signal 30 is not generated and the abnormal signal 3
Even if 0 occurs, before a predetermined period of time has passed (processing 1a)
data captures signals from detectors installed in the plant 10 and logic signals 20 representing the operating status of the plant.
The data is captured into the recording device 40 and recorded (processing 1b).

A predetermined period of time has passed since the abnormality signal 30 was generated (processing 1
a), time series data before and after the abnormal signal generation time, abnormal signal generation time data, and logic signal data 100 are sent to the event identification data generation device 110↓At the same time, plant data 101 at the same abnormal time is sent to the simulator 60. (processing 11a).

The event identification data device 110 selects data necessary for identifying an event from time-series data in a predetermined time width according to the type of the abnormal signal 30, using the abnormal signal occurrence time as a reference. Data generated from this time series data and logical signal data are combined into event identification data 120.
The event identification data device 110 is the event fixing unit 130.
(processing 1c).

The event identification unit 130 imports already learned neural network data such as weighting coefficients and threshold values obtained as a result of learning from the external storage device M140, and performs neural network processing using the event identification data. Get a certain turn.

The event identification unit 130 outputs an identification result 160 regarding this output pattern by referring to the output pattern/abnormal event correspondence data table stored in the external storage device 150 (processing 1d).

By using the identification result 160 and the abnormal part knowledge base 220, more detailed items of the abnormal part of the plant are searched in the knowledge processing bag N210 (processing 11b).
.

Finally, the identification results and abnormal items are displayed on the display terminal device 170 (processing 1]c).

For example, in the event of loss of total water supply flow rate, detailed abnormality items can be derived by determining the failed water supply pump.

(a) According to this embodiment, actual plant data can also be used as learning data, and the reliability of learning is also improved.

Therefore, in terms of identification performance, it is possible to configure an abnormal event identification device with fewer misrecognitions.

(b) Furthermore, according to this embodiment, by using knowledge processing,
Since the detailed abnormality becomes clear, it becomes easier to decide on countermeasures after an abnormal event occurs.

〔Effect of the invention〕

According to the present invention, an abnormality can be identified using data before and after the abnormality occurs, and identification accuracy can be improved (claims 1 and 3.4).

Furthermore, according to the present invention, data before and after the occurrence of an abnormality is selected at a sampling time corresponding to the type of abnormal signal and the rate of change of the detector signal, and an abnormality matching the type and rate of change can be identified. Section 2.5).

Further, according to the present invention, since logical signals are also taken in in addition to time series data, identification accuracy can be improved compared to only time series data (claim 3).

Furthermore, according to the present invention, by using the neural network for abnormality identification, the identification recognition ability of the neural network can be improved (claims 6 to 11).
).

Further, according to the present invention, by having a simulator that can simulate abnormal transient changes in a plant, it is possible to cause the neural network in the event identification device to learn (claim 7). Therefore, even if there is a change in specifications on the plant side, the event identification device can respond quickly.

Furthermore, according to the present invention, it is possible to change the sampling time width and the number of data points of data given to the input layer of a neural network during learning using a simulator (
Claim 8). Therefore, it is possible to provide the event identification device with data extracted at the optimal sampling time from time-series data of unusual events that change over time, and it is possible to reduce the rate of erroneously recognizing events.

Furthermore, according to the present invention, data regarding the number of layers of the neural network in the event identification device, the number of units that are units for processing information in each layer, and the learned neural network data corresponding thereto are stored from the external storage device 140. (Claim 9). Therefore, the configuration of the neural network used in the event identification device can be easily updated, and maintenance of the device becomes easy.

Furthermore, according to the present invention, the type of data used in the event identification device, the number of data points, the sampling time width, and the trained neural network data can be displayed on the display device in response to a user's inquiry. The confirmation power X becomes possible, and the reliability of the abnormal event identification device increases (claim 10).

Further, according to the present invention, detailed abnormal parts are clarified through the use of knowledge processing, so it becomes easy to determine countermeasures after an abnormal event occurs (claim 11).

[Brief explanation of drawings]

FIG. 1 is a block diagram of the overall configuration of the abnormal event identification device of the present invention, FIG. 2 is a flowchart explaining the operation of the device shown in FIG. 1, and FIG. 3 shows a part of the abnormal event identification device of FIG. 4 is a block diagram showing the configuration of the data importing and recording device, FIG. 4 is a flowchart explaining the operation of the device shown in FIG. 3, and FIG. 5 is a flowchart illustrating the operation of the device shown in FIG. 6 is a block diagram showing the configuration of an identification data generation device; FIG. 6 is a flowchart explaining the operation of the device in FIG. 5; FIGS. This is a diagram explaining the effect of changing the time. FIG. 8 is a block diagram showing the configuration of an event identification device and a display terminal device that constitute a part of the abnormal event identification device in FIG. 1. FIG. 9 is a flowchart explaining the operation of the device shown in FIG. 8, FIG. 10 is a configuration diagram of the output turn/abnormal event correspondence data table, and FIG. 11 is a block diagram of a modification of the abnormal event identification device shown in FIG. FIG. 12 is a flowchart illustrating the operation of the apparatus shown in FIG. 11. 10. Plant, 20. Signal from detector and plant logic signal, 30.. Abnormal signal, 31.. Simulated abnormal signal, 40.. Data acquisition/recording device, 60. Simulator, 90.. Simulated plant. Data, 100 Actual plant data, 110... Data device for event identification, 120... Data for event identification, 130... Event identification unit, 140... External storage device in which trained neural network data is recorded. , 150... External storage device recording the output pattern/abnormal event correspondence data table, 160... Data of identification results, 170...
・Display terminal device, 180, 190.200...Identification・
Learning control signal. Representative Patent Attorney Tadashi Akimoto Actual Figure Figure Cause Figure Figure Different Signal Aπ East Stomach Tsuneikichi No. B Occurrence B8 enemye*r'r'P Figure 10 Figure

Claims (1)

[Claims] 1. An abnormal event identification method using signals from a system to be identified for abnormal events before and after the abnormal signal is generated as data for abnormality identification. 2. An abnormality created by selecting signals from the system to be identified for abnormal events before and after the abnormal signal generation time at a sampling time corresponding to the type of abnormal signal, and using the selected signals as data for abnormality identification. Event identification method. 3. A recording means for recording a detection signal from a system to be identified as an abnormal event as time-series data, a detection means for detecting an abnormality occurrence time in the system to be identified for an abnormal event, and a detection means for detecting an abnormality occurrence time in the abnormal event identification target system, An abnormal event identification device comprising: means for reading out time-series data in the recording means existing at predetermined time intervals of as event identification data; and identification means for identifying an abnormal event from the event identification data. 4. Recording means for recording the detection signal from the abnormal event identification target as time series data and the logical signal from the abnormal event identification target as logical data; and the detecting means for detecting the abnormality occurrence time of the abnormal event identification target system. , means for reading out time-series data and logical data in the recording means existing in a predetermined time width before and after the abnormal signal generation time as event identification data; and the event identification data. An abnormal event identification device comprising: an identification means for identifying an abnormal event from an abnormal event; 5. The reading means includes means for specifying a predetermined time width before and after the abnormal signal occurrence time as a reference, a storage means for storing in advance a sampling time corresponding to the type of abnormal signal, and the above-mentioned reading means. means for detecting the type of generated abnormal signal and reading the corresponding sampling time from the storage means;
5. The abnormal event identification device according to claim 3, further comprising means for reading data in the recording means within the specified time width according to the sample pitch of the sampling time as event identification data. 6. The abnormal event identification device according to any one of claims 1 to 5, wherein the identification means includes a neural network. 7. The abnormal event identification device according to claim 6, wherein the neural network has a pre-learning function based on a simulator. 8. The abnormal event identification device according to claim 7, wherein during the learning, the sampling time width and the number of data points of the data to be applied to the input layer, which is the input part of the neural network, are changed in accordance with the abnormal signal. 9. The abnormal event identification device according to claim 6, wherein the number of layers and the number of units that process information in each layer regarding the neural network are given from an external memory. 10. The abnormal event identification device according to claim 6, having an inquiry function capable of displaying the type of data used in the neural network, the number of data points, the sampling time width, and the learned neural network data in response to a user's inquiry. 11. The abnormal event according to any one of claims 3 to 10, wherein the abnormal event has a function of determining further detailed abnormal items by performing knowledge processing using a database in which knowledge corresponding to abnormal events of a plant or machine is registered. Identification device.