JPH0652474A - Plant monitor device - Google Patents

Abstract

PURPOSE:To monitor the starting of an object plant to be monitored and the state of stationary operation without any experiential preprocessing by a skilled person and to detect deviation in relation among parameters which is seen in the beginning of abnormality from that in normal operation without any experiential judgment by a skilled person. CONSTITUTION:A time-series signal showing the operation state of the plant 11 is measured by a signal measuring means 12 and converted by a signal input means 13 into a digital signal, which is inputted to a signal processing means 14. The signal processing means 14 inputs and processes the time-series data in constant sampling cycles. A normal pattern learning means 31 learns at each timing point by a self-organization type neural network and generates a time-series variation pattern of ignition neurons in a two-dimensional plane. A normalcy/abnormality deciding means 32 compares the time-series variation pattern of ignition neurons obtained from the data at the time of monitoring with the pattern of normal operation obtained by the normal pattern learning means 31 and displays the result at a display means 19.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a plant monitoring device for monitoring the operating condition of a plant.

[0002]

2. Description of the Related Art Conventionally, a plant monitoring device is constructed as shown in FIG. In the figure, 11 is a plant to be monitored, and one or a plurality of time-series signals indicating its operating state are measured by the signal measuring means 12 and taken out as an analog signal. The signal measured by the signal measuring unit 12 is converted into a digital signal by the signal input unit 13 and input to the signal processing unit 14 in the monitoring apparatus body. The signal processing means 14 takes in time-series data at a certain sampling period, and performs blocking, statistical processing, data compression processing, etc. associated with the tendency of data change. The signal processed by the signal processing means 14 is
The signal is recorded by the signal recording means 15 and is also monitored by the monitoring means 1.
6 to be processed.

The monitoring means 16 comprises a signal analyzing means 17 and a normal / abnormal determining means 18, and a signal processing means 14
The signal processed by is analyzed and normality / abnormality is discriminated. The signal analysis unit 17 performs statistical processing and normal distribution determination on the data processed by the signal processing unit 14. The normality / abnormality determining means 18 compares the analysis result of the signal analyzing means 17 with a normal reference value which is empirically determined based on the normal data in advance, and determines normality / abnormality. The discrimination result of the normal / abnormal discriminating means 18 is sent to the display means 19 to be displayed and is also sent to the printing means 20 to be printed. Reference numeral 21 is an operation unit that serves as an interface when the user specifies the above-described measurement, processing, determination, display, and printing functions.

In the above configuration, the time-series data in a certain operation mode of the plant 11 measured by the signal measuring means 12 is processed by the signal processing means 1 via the signal input means 13.
Enter in 4. The signal processing means 14 is the signal measuring means 12
The time series data from is collected as a normal reference data group, and for each of these time series data, the time zone in which the same state continues is summarized as one block, and the average value of the data group during this time is the block. Compressed time-series data is created with the representative value of. And the signal analysis means 1
7, the normal distribution of the normal reference data is checked for each block, and the normal reference value is calculated by two types of statistical methods depending on the presence or absence of the property. Normal / abnormal determination means 18
Compares the normal reference value with the representative value of each block of the time-series data to be verified, and if there is even one abnormality in all blocks, it is determined that the operation is abnormal. The plant 11 is monitored by such a method.

[0005]

The above-mentioned conventional plant monitoring device empirically blocks the time-series data according to its changing tendency and compares the data at the same timing in a certain plant operation mode. Knowledge and labor of experts are required for this block formation.

Further, although it is possible to detect anomalies for each block of time series data and for each parameter, it is not possible to compare the patterns of the overall tendency of changes of a plurality of parameters, so that experience of an expert is required for early detection of anomalies. is there.

The present invention has been made in consideration of the above situation, and the start-up operation from the start of the monitored plant to the steady operation, or the state of the steady operation is monitored without empirical pretreatment by a skilled person. In addition, it is possible to detect the deviation of the relationship between each parameter seen at the initial stage of abnormality from the normal state by learning the change tendency of a plurality of parameter sets without the need for empirical judgment by a skilled plant. An object is to provide a monitoring device.

[0008]

A plant monitoring apparatus according to the present invention comprises a signal measuring means for measuring a signal from a plant to be monitored, and a signal input for inputting the signal measured by the measuring means to the apparatus body. Means, a signal processing means for compressing the signal input by this means, a signal recording means for recording the signal processed by this means, and a normal time series signal pattern processed by the signal processing means as 1 A normal pattern learning means for learning without teacher data in a network composed of a two-dimensional input layer and a two-dimensional output layer, and a normal plant from the output signal of the signal processing means based on the pattern learned by the normal pattern learning means.・ Displays the normal / abnormal judgment means for judging abnormality, the operation means for interfacing with the user, and the judgment result of the normal / abnormal judgment means. Characterized in that a display unit.

[0009]

The signal processing means inputs a plurality of time series data sampled at regular intervals during normal operation to the normal pattern learning means provided in the monitoring means. The normal pattern learning means learns with a self-organizing neural network at each timing, and creates a time-series change pattern of firing neurons in a two-dimensional plane.

The normal / abnormal discriminating means compares the time-series change pattern of the firing neuron obtained by sequentially inputting the data at the time of monitoring with the normal pattern obtained by the normal pattern learning means to make normal Or not. The determination result of the normal / abnormal determination means is sent to the display means and displayed.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the construction of a plant monitoring apparatus according to an embodiment of the present invention. The plant 11, the signal measuring means 12, the signal input means 13, the signal processing means 14, the signal recording means 15, the display means 19, the printing means 20 and the operating means 21 in FIG.
Since it is the same as the conventional device shown in FIG.

The data processed by the signal processing means 14 is sent to the monitoring means 30. This monitoring means 30
Is a normal pattern learning unit 31 that learns a pattern of normal time-series data after processing by the signal processing unit 14 and a normal / abnormal judgment that uses the learning result.
It is configured by the abnormality determination means 32.

The normal pattern learning means 31 learns a time-series change pattern of sampling data (learning data) 33 sent from the signal processing means 14 when the plant 11 to be monitored is normal. The normality / abnormality determination means 32 determines whether or not the determination data 34, which is measured and processed at regular intervals during monitoring and is input from the signal processing means 14, is normal.

FIGS. 2A and 2B are explanatory views of the normal pattern learning means 31 and have a two-layer neural network structure composed of a one-dimensional input layer 41 and a two-dimensional output layer 42. The input is a set of data of a plurality of parameters for every 5 minutes standardized in the range of 0 to 1, and there are a total of 13 sets for an operating time of 60 minutes. These input data are assigned to one of the neurons in the output layer 42 by repeatedly learning them by a self organizing feature mapping method without supervision, and a time series change pattern of the firing neuron is obtained. That is, the normal pattern learning means 31 is self-organizing at each timing with respect to the learning data 33 sent from the signal processing means 14, that is, a plurality of time series data sampled at regular intervals during normal operation. By learning with a neural network, a time-series change pattern of firing neurons is created in a two-dimensional plane.

The above 13 sets of data will be described in more detail. As shown in FIG. 2A, the signal processing unit 14 outputs a total of 13 time-series data for each of the three types A, B, and C, for example, every 5 minutes from 0 to 60 minutes. The three types of data (A, B, C) at each time are set as one set,
A total of 13 data can be created. Normal pattern learning means 3
1 inputs these 13 pieces of data in order from the input layer 41 as shown in FIG. 2B, and for each data, the output layer 4
From the 7 × 7 output neurons in 2, one corresponding neuron is determined.

The method of selecting the corresponding output neuron is performed as follows. Consider the three input data as a three-dimensional vector. Also, 4 in the output layer 42
Similarly, the data of the weight values of the three links from each of the nine output neurons to the three input neurons of the input layer 41 is also regarded as a three-dimensional vector. These are given initial values as random numbers, and for each of the 13 input vectors, 4
The distances from the nine weight vectors are calculated, and the one having the smallest distance is determined as the corresponding output neuron.
It is called firing that this distance is determined and selected as the minimum, and the selected neuron is called a firing neuron.

Accordingly, among the 13 input vectors, if there are close ones, the corresponding output neurons may match. In this way, the firing pattern of the output neuron shown in FIG. 3 represents a state in which 13 or less corresponding output neurons are determined and fired in the order of input.

In FIG. 3, P1 is a time-series change pattern of firing neurons, which is a learning result of normal operation data, and P1
2 is the change pattern of the firing neuron obtained by making the learned neural network recognize the data during abnormal operation. This is the change pattern P2
Indicates that the change pattern P2 is out of the normal change pattern.

The normal / abnormal discriminating means 32 has a normal pattern obtained by the normal pattern learning means 31 and a time-series change pattern of firing neurons obtained by sequentially inputting discrimination data 34 for monitoring. Is compared to determine whether or not it is normal. The result discriminated by the normal / abnormal discriminating means 32 is displayed on the display means 19
Is displayed by the printer, and is printed by the printing unit 20 as necessary.

[0021]

As described above in detail, according to the present invention, the start-up operation from the start-up of the monitored plant to the steady operation, or the state of the steady operation is monitored without empirical pretreatment by a skilled person. be able to. Further, by learning the change tendency of a plurality of parameter sets, it is possible to detect the deviation of the relationship between the parameters, which is observed at the initial stage of abnormality, from the normal state, without empirical judgment by a skilled person.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a plant monitoring apparatus according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of normal pattern learning means in the same embodiment.

FIG. 3 is a diagram showing an example of a normal / abnormal determination means in the embodiment.

FIG. 4 is a diagram showing a configuration of a conventional plant monitoring device.

[Explanation of symbols]

11 ... Plant, 12 ... Signal measuring means,
13 ... Signal input means, 14 ... Signal processing means, 15
... signal recording means, 16 ... monitoring means, 17 ... signal analyzing means, 18 ... normality / abnormality determining means, 19 ... display means, 20 ... printing means, 21 ... operating means,
30 ... Monitoring means, 31 ... Normal pattern learning means, 32 ... Normal / abnormality discrimination means, 33 ... Learning data 34 ... Discrimination data, 41 ... Input layer, 4
2 ... Output layer.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor, Seiji Yaguchi 1-1-1, Wadazaki-cho, Hyogo-ku, Kobe-shi, Hyogo Mitsubishi Heavy Industries, Ltd. Kobe Shipyard (72) Hiroyuki Nakayama, Niihama, Arai-cho, Takasago-shi, Hyogo 2-1-1, Mitsubishi Heavy Industries, Ltd. Takasago Research Center

Claims (1)

[Claims] 1. A signal measuring unit for measuring a signal from a plant to be monitored, a signal input unit for inputting a signal measured by the measuring unit to a main body of the apparatus, and a signal input by this unit is compressed. A signal processing means, a signal recording means for recording the signal processed by this means, and a network consisting of a one-dimensional input layer and a two-dimensional output layer for the time-series signal pattern in a normal time processed by the signal processing means. A normal pattern learning means for learning without teacher data, a normal / abnormal judgment means for judging a normal / abnormal of a plant from an output signal of the signal processing means based on a pattern learned by the normal pattern learning means, and a user And a display unit for displaying the discrimination result of the normality / abnormality discrimination unit. Vision device.