JPH06109498A - Detector for unsteady and abnormal state - Google Patents

Abstract

PURPOSE:To detect the unsteady and abnormal state of a plant accurately even if noise is included in time series data or irregular disturbance is generated. CONSTITUTION:The history of the observation signal of a plant 2 is kept in a storage 3 for storing as time-series data. The behavior of the plant 2 is estimated by two antoregressive models 4 and 5. Both antoregressive models 4 and 5 are constituted by a neural network in three-layer structure of input layer, middle layer, and output layer. Also, the number of middle layer units of antoregressive models 4 and 5 differs. In the case of the neural network where the number of middle layer units differs, a nearly equal estimation value is obtained in normal state. On the other hand, a different estimation value is obtained in abnormal state. Therefore, by comparing the estimation values of antoregressive models 4 and 5 with an arithmetic unit 6, the abnormal state of the plant 2 can be detected simply and readily.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for detecting unsteady and abnormal states, and in particular, it monitors a plant in real time in a power plant, a chemical plant or the like which is affected by complicated noise, irregular disturbances. The present invention relates to an unsteady state and abnormal state detection device suitable for use in the case.

[0002]

2. Description of the Related Art For the purpose of predicting the state of a plant, an autoregressive model, which is a statistical method, has been used for a long time. The detection of unsteady state and abnormal state using the autoregressive model is performed by monitoring the prediction error (hereinafter referred to as the residual) between the real time series data and the autoregressive model. On-line processing is desired for detection of unsteady and abnormal states, except when enormous statistical calculations are required such as system diagnosis. There is a limit to the amount of calculation in online processing, statistical processing such as variance for enormous time series data is rarely performed, and when the residual value shows a value higher than a predetermined threshold, It is judged to be steady and abnormal.

[0003]

In the conventional non-steady state and abnormal state detecting device, when the time series data obtained from the system contains a high level of noise, or when an irregular disturbance occurs, The accuracy of the model fitting certainly deteriorates, and the threshold for residuals for detecting non-steady state and abnormal state must be set to a high level as a matter of course.Therefore, there is a considerable delay in non-steady state detection. There is a problem that it will occur. The present invention has been made in consideration of the above-mentioned circumstances, and is a case where the obtained time series data is contaminated by noise and the accuracy of the model fitting and the estimation by the model is low due to irregular disturbance. Even if, however, it is an object of the present invention to provide an unsteady state and abnormal state detection device that can detect an unsteady state and a divergent abnormal state of a plant online and at an early stage.

[0004]

In order to solve the above-mentioned problems, an unsteady state and abnormal state detecting apparatus according to the present invention comprises:
A neural network having a storage device that stores the history of observation signals of a plant to be monitored as time-series data, a hierarchical type having an input / output layer and an arbitrary number of intermediate layers, and different intermediate layer structures. Each of which is configured with a plurality of autoregressive models for estimating the behavior of the plant by inputting the time series data from the storage device, and mutually comparing the estimated values of the behavior of the plant output from each of these autoregressive models, And an arithmetic unit for determining an unsteady state and a divergent abnormal state.

[0005]

The unsteady state and abnormal state detecting apparatus according to the present invention uses a plurality of autoregressive models, and the estimated values of the behavior of the plant output from these autoregressive models are compared with each other to obtain the plant. Unsteady and divergent abnormal conditions of are detected.

By the way, a plurality of neural networks having mutually different intermediate layer structures can output almost the same predicted value for the behavior of the system seen in the steady state. Therefore, it is not necessary to consider the influence on the residual error caused by the model, noise, and irregular disturbance.

On the other hand, for the unsteady state and the divergent abnormal state, the difference in the structure of the intermediate layer appears as a difference in the output, and the comparison makes it possible to detect the unsteady state and the divergent abnormal state. Become. Moreover, since the threshold value can be set to an extremely low level, it is possible to detect it earlier than in the conventional case.

The term "unsteady" as used herein means a state in which the user is unsteady, and the term "divergent" means a phenomenon involving a larger vibration than a shaking that is not normally seen.

[0009]

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

FIG. 1 shows an unsteady state and abnormal state detecting apparatus according to the present invention. This detecting apparatus 1 records the history of observation signals of a plant 2 to be monitored, such as a power plant or a chemical plant. The storage device 3 for storing as series data is provided, and the time-series data stored in the storage device 3 are two autoregressive models 4, 5 composed of neural networks having mutually different intermediate layer structures. The behavior of the plant 2 is estimated by each of the autoregressive models 4 and 5. Then, the estimated values of the behavior of the plant 2 from the respective autoregressive models 4 and 5 are compared with each other in the arithmetic unit 6, and the plant 2 is in a steady state, or an unsteady state and a divergent abnormal state (hereinafter, It is determined whether or not it is a non-steady state).

The neural network for constructing each autoregressive model 4 and 5 is basically constructed as a hierarchical type having a three-layer structure of an input layer, an output layer and an arbitrary number of intermediate layers, and the input / output layers are In addition to having linear characteristics, the middle layer has nonlinear characteristics (a sigmoid function is generally used)
Is configured as having.

When a univariate or multivariate vector is represented by a symbol x, the state quantity z (t) of the plant at a certain time t is the past state quantity {x (t-Δt) sampled at the sampling time Δt. ), ……, x (t
-MΔt)},

[Equation 1] (However, f is the input / output relationship acquired by the neural network, and ε is the residual.), Which is called an autoregressive model. In this embodiment, a structure as shown in FIG. 2 is adopted in order to construct an m-th order autoregressive model composed of m measurement values shown in Expression 1 on the neural network.

Further, each of the autoregressive models 4 and 5 in this embodiment has the same number of input / output layer units (in this case, the input layer has m units and the output layer has one unit) and different numbers of intermediate layers. It is composed of units, and the difference in the number of units in the middle layer appears as the difference in the function form expressed by the neural network. If this is expressed by an equation, the autoregressive model 4 becomes

[Equation 2] (However, ε ₁ (t) is the residual.) Further, the autoregressive model 5 is

[Equation 3] (However, ε ₂ (t) is a residual.)

In equations (2) and (3), the functions f1 and f2 of the neural networks having different intermediate layer structures are
It is obtained by time-series learning, and it is possible to output almost the same predicted value for the behavior (amplitude) of the system seen in the steady state. However, both neural networks output different prediction values when an unlearned and amplitude that is not found in the steady state is input. The prediction values are different because there is a difference in the non-linearity expressing both functions f1 and f2. FIG. 3 shows the functions f1 and f
This is a schematic representation of the relationship of 2, and it can be said that the functions f1 and f2 are functionals in the steady state.

Next, the determination of the number of input layer units will be described.

When the target time series data is linear, the neural network uses a linear autoregressive model.

[Equation 4] (However, w is a coefficient that characterizes the system.) A structure equivalent to that is acquired by learning. When the intermediate layer is omitted in the neural network using the linear unit for the input / output layer, w _i (i = 1, ..., M) in the above equation matches the coupling coefficient between the input / output layer units. Therefore,
Following the fitting of a general linear autoregressive model, AIC
The smallest order is adopted as the number of input layer units.

Here, the minimum order of AIC is as follows. That is, the AIC (Akaike Information Standard Akaike Informati) is used as a criterion for determining the structure of the autoregressive model.
onCriterion).

The definitional expression is

[Equation 5] Given in. Likelihood here means the "probability" of the model and depends on the problem.

The "AIC minimum" means that the model has a high probability from the definitional expression and that the number of parameters describing the model (the order in the conventional linear autoregressive model, the number of connected neuron units in the neural network) is small. Means And this autoregressive model
It is considered to be an optimal model with no waste.

On the other hand, when the target time-series data is non-linear, the fractal dimension is obtained and the smallest integer dimension including the fractal dimension is adopted as the number of input layer units.

Here, the fractal dimension is a generalized concept of the dimension. While the ordinary dimension takes an integer value (for example, a value that does not include a decimal point such as two dimensions), the fractal dimension has a real value (for example, two values). Value including a decimal point such as .07 dimension). In order to measure the fractal dimension of a shape, the shape must be a single simple figure (segment, circle, rectangle,
This is called a metric, such as a sphere or a cube. ) Consider covering with. When the number N of metrics required for covering has a relationship of N (r) to r ^{ν with} respect to the metric size r, this exponent ν is called a fractal dimension. If the measurement target is an integer-dimensional shape (such as a cube), the fractal dimension ν
Also takes an integer value.

When convergence of the fractal dimension is obtained, a trajectory having a structure can be reconstructed from the time series data. Highly accurate estimation is possible by learning the structure of this trajectory in a neural network. On the other hand, if the convergence of the fractal dimension cannot be obtained, the linear time series data is used. Here, the trajectory is a transition of the state (or position) of the system, and the "orbit" is one of them.

Next, the unsteady detection method will be described.

The input / output relationship f is acquired by learning the same stationary time series data for two neural networks having different numbers of intermediate layer units. By learning the same stationary time series data, it is possible to obtain almost the same estimated value in the stationary state even with autoregressive models having different structures. However, in the non-steady state, the difference in structure causes a difference in the state prediction ability, so that each autoregressive model outputs different estimated values. Therefore, a non-steady state can be detected by monitoring the output difference of autoregressive models (neural networks) having different structures.

Specifically, a threshold value is set for the output difference of the autoregressive model (neural network) having a different structure, and when the output difference exceeds the threshold value, it is judged as an unsteady state or an abnormal state. As shown in FIG. 2, an autoregressive model is constructed on the neural network. Input the past state quantities {x (t-Δt), ..., x (t-mΔt)} sampled at the sampling time Δt to each input layer unit. According to this method, the time is directly input to the neural network. Input and output to the autoregressive model {x (t-Δt),
.., x (tm-t)} does not need to be standardized to the output range of the intermediate layer (generally, [0, 1] for a sigmoid function). Also, the number of input layer units is optimized according to the characteristics of the target time series by the AIC and the fractal dimension.

Next, a specific example will be described.

We compare the outputs for stationary unlearned and non-stationary data of two neural networks with different hidden layer structures. FIG. 4 shows the time series data used at that time. Steady and non-steady behaviors are recorded in this time series data. FIG. 5 and FIG. 6 show the absolute values of residuals when time-series data is passed through neural networks having different structures. The absolute value of the output difference between both neural networks is shown in FIG.

In FIG. 5, the input / output layer is a linear unit and the intermediate layer is a non-linear unit, and the number of units is the absolute value of the residual error in the case of the neural network of the input layer 20, the intermediate layer 1 and the output layer 1. 6 is a linear unit in the input / output layer and a non-linear unit in the intermediate layer, and the number of units is the absolute value of the residual in the case of the neural network of the input layer 20, the intermediate layer 1 and the output layer 1. is there.

FIG. 8 is an enlarged view of the input time series data to the neural network in the case of divergent abnormality, and FIG.
Indicates the prediction value of the neural network for this time series data. In FIG. 9, + is the number of intermediate layer units 1
The predicted value in the case of, and the solid line is the predicted value in the case of 16 hidden layer units.

As is apparent from FIG. 9, it can be seen that neural networks having different structures output different prediction values for non-stationary states. This is because the input / output relationship of the neural network 7 acquired by learning is expressed by the sum of the nonlinear functions of the intermediate layer units.

FIG. 10 and FIG. 11 are enlarged views of this. That is, FIG. 10 shows the predicted value of the neural network in the steady state, and FIG.
Indicates the prediction value of the neural network in the non-steady state. In both figures, the thick line is the predicted value when the number of intermediate layer units is 1, and the thin line is the predicted value when the number of intermediate layer units is 16.

In the steady state, the predicted values are almost the same as shown in FIG. 10 and cannot be distinguished, but in the non-steady state, both are as shown in FIG. Since the predicted values are deviated, it becomes possible to detect the abnormality in the portion indicated by the arrow in FIG. This detection is shown in FIG.
Is performed by the arithmetic unit 6 of.

FIG. 12 shows the flow of processing in the arithmetic unit 6.

The computing device 6 first compares the predicted values of the neural networks of the two autoregressive models 4 and 5 in step S1 and detects the maximum difference between the predicted values in step S2. Then, in step S2, it is determined whether or not the maximum value of the difference between the predicted values exceeds the threshold value. If not, it is determined in step S4 that it is in the normal state.
In step S5, it is determined that the state is abnormal.

Thus, it is shown that the outputs of both neural networks are almost equal in the steady state. In other words, by comparing the outputs of autoregressive models with different structures, the difference in the structure of the middle layer appears as a difference in the output for models, noise, and irregular disturbances. It is possible to detect the non-steady state by comparing the outputs. Moreover, the output difference between both neural networks in the steady state is extremely small, and the threshold value for the output difference of the neural network used for non-steady state detection can be set to an extremely low level. From this, unsteady detection becomes possible earlier than the conventional method.

In FIG. 5, when 1.033, which is 1.2 times the maximum residual 0.861 in the steady state, is used as the threshold for detecting the non-steady state, the non-steady state is detected in 338.125 seconds. . In FIG. 7, when 0.016, which is 1.2 times the maximum output difference 0.080 of the neural network in the steady state, is used as the threshold for detecting the non-steady state, the non-steady state is detected in 318.250 seconds. . From this, it is found that this method is effective for early detection of an online steady state.

[0037]

As described above, according to the present invention,
The obtained time series data is polluted by noise,
Moreover, even when the accuracy of the model fitting and the estimation by the model is low due to the irregular disturbance, the unsteady state and abnormal state of the plant can be detected online and early.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram showing an unsteady state and abnormal state detecting apparatus according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing construction of an autoregressive model using a neural network.

FIG. 3 is a graph schematically showing a relationship between functions of two neural networks having different intermediate layer structures.

FIG. 4 is a graph showing an example of time series data passed through a neural network.

FIG. 5 is a graph showing absolute values of residuals of a neural network having one hidden layer unit.

FIG. 6 is a graph showing absolute values of residuals of a neural network having 16 hidden layer units.

FIG. 7 is a graph showing the absolute value of the difference between the estimated values of two neural networks having the number of hidden layers of 1 and 16.

FIG. 8 is a graph showing a main part of input time series data to a neural network in a divergent abnormal state.

9 is a graph showing predicted values for the data in FIG. 8 of two neural networks having the number of hidden layers of 1 and 16, respectively.

FIG. 10 is a graph showing predicted values in a steady state of two neural networks having 1 and 16 hidden layers.

11 is a graph showing prediction values in a non-steady state of the neural network similar to FIG.

FIG. 12 is a flowchart showing a processing procedure in the arithmetic device.

[Explanation of symbols]

 1 Detection Device 2 Plant 2 Storage Device 4,5 Autoregressive Model 6 Computing Device

Claims (1)

[Claims] 1. A storage device for storing the history of observation signals of a plant to be monitored as time-series data, and a hierarchical structure having an input / output layer and an arbitrary number of intermediate layers, and the structures of the intermediate layers are mutually mutual. A plurality of autoregressive models each of which is composed of different neutral networks and estimates the behavior of the plant by inputting the time series data from the storage device, and an estimated value of the behavior of the plant output from each of these autoregressive models. An unsteady state and abnormal state detection device, comprising: a computing device that compares the states with each other to determine an unsteady state and a divergent abnormal state.