JPH10198420A - Method and device for diagnosing abnormality - Google Patents

Abstract

PROBLEM TO BE SOLVED: To estimate the factor of abnormality by learning the relation of measured value in abnormal state through an abnormal state learning part and comparing it with the relation of cause and effect in normal state when a process is turned to the abnormal state. SOLUTION: At a normal state learning part 30, the relation of cause and effect between normal data stored in a data storage part 20 is learnt by a neural network. The weight coefficient of neural network is reflected with the relation of cause and effect provided by learning. When any abnormality is discriminated by an abnormality detection part 40, the measured value in the abnormal state is sent to an abnormal state learning part 50, and the relation of cause and effect between the measured values in the abnormal state is learnt. The weight coefficient of neural network is reflected with the learnt result similarly to the normal state learning part 30. The weight coefficients provided by the normal state learning part 30 and the abnormal state learning part 50 are sent to an abnormality factor estimating part 60 and both the coefficients are compared and analyzed. Then, the abnormality factor is narrowed by analyzing the relation of cause and effect between the measured values in the normal and abnormal states.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an abnormality diagnosis method and apparatus.

[0002]

2. Description of the Related Art An anomaly detection method using a neural network includes "anomaly detection based on a functional relationship between process variables by a self-associative neural network" (Chemical Engineering Transactions, Vol. 22, No. 4, p846-p853). ,
1996). In known examples,
Using a three-layer or five-layer neural network, when the process is in a normal state, the measured values at each time of the measured time series data are input to the input layer, and the output layer is learned so as to match the input layer. . When the measurement data is input to the neural network in which the relationship between the measurement values in the normal state has been learned in this manner, if the process state is normal, the output value will be the same as the input value, and if the process state is abnormal, the input value will be different from the input value. Will output different values. Therefore, a fixed reference value is provided for the difference between the input value and the output value, and if the difference is within the reference value,
It is possible to determine the state of the process such that the state of the process is normal, and if it exceeds the reference value, it is abnormal.

[0003]

As described above, according to the prior art, it is possible to determine whether or not a process is in a normal state, and to detect an abnormality. However, when the process becomes abnormal, information for identifying the cause of the abnormality cannot be obtained.

[0004]

Means for Solving the Problems To solve the above problems, an abnormal state learning unit for learning the relationship between the measured values of the abnormal state, and a weight coefficient and an abnormal state of the neural network that has learned the normal state have been learned. It has an abnormal cause estimating unit that compares the weighting factors and estimates the abnormal cause.

[0005]

FIG. 1 shows an embodiment of the present invention.

This embodiment is for detecting an abnormality of the emergency diesel engine generator and estimating the cause of the abnormality when the abnormality occurs.

[0007] The emergency diesel engine generator starts when a power failure occurs and the supply of commercial power is stopped. That is, since the emergency diesel engine generator is not normally operated, a test operation is periodically performed to perform maintenance and inspection. In this embodiment, a state quantity such as a generator voltage and a field current is measured from a diesel engine generator during a test operation, and abnormality detection and abnormality cause estimation are performed from the measured data.

This embodiment comprises an emergency diesel engine generator 1, a communication line 2, and a monitoring center 3. The monitoring center 3 comprises an abnormality diagnosis device 4 and an operation terminal 70. The abnormality diagnosis device 4 includes a data preprocessing unit 10, a data storage unit 20, a normal state learning unit 30, an abnormality detection unit 40, an abnormal state learning unit 50, and an abnormality cause estimating unit 60.

Next, an outline of the operation of the present embodiment will be described. This embodiment has a learning mode and an abnormality detection mode.

[0010] The learning mode is a preparation stage for operating the abnormality diagnosis device, and is a mode in which the emergency diesel engine generator 1 learns the accumulation of data in a normal state and its causal relationship. In the learning mode, the measurement data of the diesel engine generator 1 is checked by the data pre-processing unit 10 for any abnormal data due to measurement and communication, such as missing data and measurement range over. If there is an abnormality, it is stored in the data storage unit 20 as abnormal data, and if there is no abnormality, it is stored as normal data. However, the data storage unit 20
Is not necessarily guaranteed to be data in a normal state. Therefore, if the data stored as normal data is found to be abnormal due to inspection work or the like, it is stored as abnormal data.

In the normal state learning unit 30, the data storage unit 2
A causal relationship between normal data stored in 0 is learned by a neural network. The causal relationship obtained by learning is reflected in the weighting factor of the neural network.

Next, the abnormality detection mode will be described.
The abnormality detection mode is a mode in which abnormality is diagnosed after sufficient learning has been performed in the learning mode.

In the abnormality diagnosis mode, the measured data is checked by the data pre-processing unit 10 for abnormal data by measurement and communication as in the learning mode.
Sent to The abnormality detection unit 40 inputs the measurement data to the neural network having the weight coefficient obtained by the normal state learning unit in the learning mode, and calculates the output value of the neural network. The output value is compared with the normal state, and if the deviation is equal to or smaller than the reference value, it is determined that the output is normal, and if the deviation is larger than the reference value, it is determined that the output is abnormal. The determination result is sent to the operation terminal 70, and the determination result is displayed on a monitor screen of the operation terminal 70.

When the abnormality detection unit 40 determines that an abnormality has occurred,
The measured value of the abnormal state is sent to the abnormal state learning unit 50, and the causal relationship between the measured values of the abnormal state is learned. The learning result is reflected on the weight coefficient of the neural network as in the normal state learning unit 30. The weighting factors obtained by the normal state learning unit 30 and the abnormal state learning unit 50 are sent to the abnormality cause estimating unit 60, and both are compared and analyzed. Then, the cause of the abnormality is narrowed down by analyzing the causal relationship between the measurement values of the normal state and the abnormal state. The cause of the abnormality narrowed down by the analysis is sent to the operation terminal 70 and displayed on the monitor of the operation terminal 70.

Next, the details of this embodiment will be described.

The emergency diesel engine generator 1 includes:
Sensors for measuring the generator engine speed, the generator voltage, the generator current, the generator field current, the starting battery total voltage, the cooling water pressure, and the battery fluid temperature are provided. These sensors can set the measurement range, and generate a "measurement range over" signal when the measured value does not fall within the measurement range. Even if the sensor is not a sensor whose measurement range can be set in this way, an area to be used as a measurement value is set in advance, and a "measurement range over" signal is generated when the measurement value from the sensor exceeds this set area. The device for performing the operation may be provided in the sensor.

These measurement data are measured by applying a start command from the operation terminal 70 of the monitoring center 3 and designating a collection time and a collection interval. Therefore, for example, when the emergency generator is started up and 60 seconds at intervals of 0.1 second, 600
Point, when stopping at 60 seconds at 0.5 second intervals, 120 points,
Assuming 300 seconds at 30-second intervals during normal operation, the number of measurement data to be collected is determined, such as 10 points. Further, the measurement data transferred from the emergency diesel engine generator 1 is checked by the data pre-processing unit 10 for the presence or absence of an abnormality and loss of the measurement data. The determination of abnormalities in the measurement data is based on whether or not a signal of “measurement range over” is added to the measurement data. Since the number is determined, it can be determined by comparing with the number of measurement data actually transferred.
In this way, it is determined whether or not there is an abnormality, and the measurement data determined to be abnormal is stored in the abnormal data storage file of the data storage unit 20. On the other hand, when it is not determined to be abnormal, the data is stored in the normal data storage file of the data storage unit 20.

In the normal state learning unit 30, the data storage unit 2
The normal data in 0 is learned, and the weight coefficient of the neural network is determined. FIG. 2 shows the neural network used in this embodiment. An input layer 7, an intermediate layer 5, and an output layer 7. The feature of this neural network is that it has a bottleneck structure in which the number of input layers and the number of output layers are symmetrical and the number of intermediate layers is smaller than the number of input layers. The time series measurement data (n items) of the plant is input to such a neural network, and the backproperty is set so that the i-th (i = 1 to n) input is equal to the i-th (i = 1 to n) output. It is known that a causal relationship between measurement data can be learned by learning by a gating method (see the above-mentioned known example). In this embodiment, the above-described seven items of time-series data measured at the time of starting the emergency diesel engine generator are provided as teacher data.
That is, as shown in FIG. 3, the measured values of seven items for each time were set as one set of input, and 100 sets of measured values were used as teacher data. Learning was performed by back propagation so that the input value and the output value were equal, as in the above-described known example. In the present embodiment, data up to 60 seconds after startup is measured and stored in the data storage unit 20. However, since the emergency diesel engine generator in question reaches a steady state approximately 7 seconds after startup, The measured values in the subsequent 10 to 60 seconds are substantially constant. Since the learning effect is small even if the same data is used as the teacher data, the teacher data is 100 points of data from 0 to 10 seconds. Further, depending on the diesel engine generator, the time required to reach the steady state is even shorter, and the teacher data may be insufficient. In such a case, test operation data on another day may be added and used as the teacher data.

Next, the abnormality detection mode will be described.
The abnormality detection mode is a mode for actually performing abnormality diagnosis.
The measurement data to be determined is sent to the abnormality detection unit 40 after the data preprocessing unit confirms that there is no missing data or the like.

The abnormality detection unit 40 inputs measurement data for each time to the learned neural network.
Since the weighting factor of the neural network stores the causal relationship between the measurement values in the normal state, the i-th (i = 1 to 7) output value is the i-th (i = 1 to i) in the case of normal measurement data.
7) are substantially equal to the input values. In the case of an abnormal state, the output value is different from the input value. Therefore, in the present embodiment, an abnormality is detected using the mean square error of the first to seventh input / output values as an evaluation criterion. That is, the ith (i = 1 to
7) If the mean square error of the difference between the input value and the output value was equal to or less than the reference value, it was judged as normal, and if it exceeded the reference value, it was judged as abnormal. The measurement data determined to be normal or abnormal by the abnormality detection unit 40 is stored in the normal data file and the abnormal data file of the data storage unit. Further, the determination result is displayed on a monitor screen of the operation terminal 70 and transmitted to the operator.

The abnormal state learning section 50 learns the causal relationship between the measured values in the abnormal state using the data determined to be abnormal by the measurement abnormality detecting section 40. This learning method is exactly the same as that of the normal state learning unit 30.

The abnormal cause estimating section 60 estimates the abnormal cause from the weighting factors obtained by the normal state learning section 30 and the abnormal state learning section 50. FIG. 4 shows an outline of the algorithm.

In the first step, the weight coefficients of the normal state and the abnormal state are compared, and a weight having a large deviation is extracted.
In this embodiment, a coefficient whose deviation is larger than 5% is extracted based on the normal state. Next, an abnormal measurement item is narrowed down from the extracted weight coefficient. Now, the weight coefficient between the input layer node and the intermediate layer node is represented by w1 (i, j) (i = 1 to 7, j =
1-5) The weight coefficient between the intermediate layer node and the output layer node is w
2 (j, k) (j = 1 to 5, k = 1 to 7). In this embodiment, if a combination of w1 (i, j) and w2 (j, k) having the same value of j is found from the extracted weighting factors, it is determined that the causal relationship between the item i and the item k is abnormal. And
Item i and item k were extracted as abnormal measurement items. For example, the weighting factors whose deviations are larger than 5% in the comparison process are w1 (1,2), w1 (3,1), w1 (1,5), w1
In the case of 2 (2,3), w2 (3,2), w2 (3,1), the item w1 (1,2), w2 (2,2) has j = 2. It is determined that the causal relationship between 1 and item 3 has become abnormal.

After narrowing down the abnormality measurement items,
Identify specific abnormalities. In the present embodiment, the determination is performed using the correspondence table between the measurement item and the cause of the abnormality. this is,
The table summarizes measurement items that indicate abnormal values when a certain part becomes abnormal. An example is shown in FIG. For example, if the extracted abnormality measurement items are the field voltage and the generator voltage, it is determined from this table that the cause when the correlation between the field voltage and the generator voltage is abnormal is “AVR abnormality”. .

When estimating the cause of an abnormality from an abnormality measurement item, an estimation rule may be created and the cause of the abnormality may be estimated by knowledge processing.

The cause of the abnormality thus estimated is displayed on the monitor screen of the operation terminal 70. If the process of the estimation is necessary, it can be similarly displayed on the monitor screen of the operation terminal 70.

[0027]

According to the present invention, when the process enters an abnormal state, the abnormal state learning unit learns the relationship between the measured values of the abnormal state,
The cause of the abnormality can be estimated by comparing with the causal relationship in the normal state.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the present invention.

FIG. 2 is an explanatory diagram schematically showing a neural network.

FIG. 3 is an explanatory diagram illustrating an example of teacher data (input).

FIG. 4 is a flowchart illustrating an algorithm of a cause estimating unit.

FIG. 5 is an explanatory diagram showing an example of a relation table between a cause of abnormality and a measured value.

[Explanation of symbols]

1. Emergency diesel engine generator, 2. Communication line,
3 monitoring center, 4 abnormality diagnosis device, 10 data preprocessing unit, 20 data storage unit, 30 normal state learning unit, 4
0: Abnormality detecting unit, 50: Abnormal state learning unit, 60: Abnormal cause estimating unit, 70: Operation terminal.

Continued on the front page (72) Inventor Masahiro Yoshioka 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture Inside the Omika Plant, Hitachi, Ltd. (72) Inventor Masanori Kanaga 5-2-1 Omika-cho, Hitachi City, Ibaraki Prefecture (72) Inventor Satoshi Goto 5-2-1, Omika-cho, Hitachi City, Ibaraki Prefecture In-house Omika Plant, Hitachi, Ltd.

Claims (3)

[Claims] 1. An abnormality diagnosis method for detecting an abnormality of an apparatus from measurement data of the apparatus and estimating a cause of the abnormality, wherein a neural network model which inputs and outputs measurement values measured from the equipment in a normal state in advance. A normal state learning unit that learns the relationship between the measurement values in the normal state by the neural network, and an abnormality detection unit that inputs the measurement value measured at the time of determination and compares the obtained output with the output in the normal state. When the abnormality is determined to be abnormal by the abnormality detection unit, the relationship of the abnormal data is learned by the abnormal state learning unit, and the weight factor of the neural network at the time of abnormality is compared with the weight coefficient of normal to determine the cause of the abnormality. An abnormality diagnosis method comprising an abnormality cause estimating unit for estimating the abnormality. 2. An abnormality diagnosis apparatus for detecting an abnormality of an apparatus from measurement data of the equipment and estimating a cause of the abnormality, using a neural network model for inputting and outputting measured values measured from the equipment in a normal state in advance. Normal state learning means for learning the relationship between the measurement values in the normal state by the neural network, and abnormality detection means for inputting the measurement value measured at the time of determination, and comparing the obtained output with the output in the normal state. When the abnormality is determined to be abnormal by the abnormality detecting unit, the relation of the abnormal data is learned by the abnormal state learning means, and the cause of the abnormality is estimated by comparing the weight coefficient of the neural network at the time of abnormality with the weight coefficient of the normal state. An abnormality diagnosis apparatus characterized by having an abnormality cause estimating means for performing the operation. 3. A diesel engine power generation system comprising the abnormality diagnosis device according to claim 2, wherein a measured state quantity of the system is input to the abnormality diagnosis device to perform abnormality detection and abnormality diagnosis.