JP3659723B2 - State change alarm device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a system state change alarm device , and more particularly, to a state change alarm device applying a deterministic nonlinear system (chaos) theory.
[0002]
[Prior art]
Conventionally, abnormality monitoring technology has been required in various fields, and it is required to accurately and promptly determine whether a device to be monitored is in an abnormal state or a normal state depending on the situation. Yes.
[0003]
Usually, in order to make such a determination, the state of the monitored object is represented by various parameters such as temperature, humidity, speed, etc., and these parameters have exceeded a predetermined number of times or a predetermined time upper limit value or lower limit value. In this case, it is determined that the monitoring target is in an abnormal state. In recent years, AI (artificial intelligence) and fuzzy reasoning are used to determine abnormalities and normal states by rules generated based on empirical rules and accumulated data.
[0004]
As described above, in any of the above-described techniques, as shown in FIG. 3, the monitoring target is abnormal by approximating the normal state from the past data and dividing the region into the abnormal region / normal region. Or whether it is normal.
[0005]
[Problems to be solved by the invention]
However, when the system is dynamic, accurate determination is difficult only by the bisection method using the abnormal region / normal region as described above. For this reason, although the normal area is currently set with a margin, the normal area is eventually determined as a fixed area.
[0006]
Therefore, it is difficult to change the real-time normal area / abnormal area in accordance with the state of the parameter or the like. Further, as described above, since the normal area must be set with a margin, there is a possibility of erroneous determination that an abnormal state exists even in the area determined to be normal.
[0007]
Further, for example, even when the monitoring target is changing from the normal state to the abnormal state, the conventional method can only determine that the current state is the normal state. Therefore, it cannot be determined whether the current state is normal or the state is approaching an abnormal state or is stable in the normal state.
[0008]
In addition, the state change of the monitoring target is predicted in the near future by a linearization method such as the least square method or the AR method, and the trend change is predicted, and the normal or abnormal state of the monitoring target is predicted from the state change. It is also possible to issue a judgment and issue an alarm. However, the least square method and the AR method are not easy to set parameters for identifying the system, and the parameters once set require a lot of time to be reset if the dynamics of the system changes. There is.
[0009]
The present invention has been made under the above background, to provide a state change alarm system and automatically identified can be made unnecessary resetting of parameters using only the time-series data the dynamics of the system Is an issue.
[0010]
[Means for Solving the Problems]
In order to solve the above problems, the present invention provides a storage unit for time series data of an observation system,
A time vector after a predetermined step is generated by chaos inference based on the behavior of the attractor generated in the reconstruction space by generating a data vector using the time series data obtained from the storage unit as a parameter, and embedding the data vector A parameter optimization that performs parameter optimization processing when the prediction error that increases the prediction value by comparing the prediction value obtained from this prediction unit with the observation value in the observation system Part, the predicted value and the observation value of the observation system measured after the predetermined step elapses are extracted and input, and the maximum Lyapunov exponent and the fractal dimension based on the attractor using the parameter from the parameter optimization unit And an identification unit for identifying system dynamics composed of predicted values and observation values, and a system from this identification unit. A recording unit for recording the dynamics, a change monitoring unit for detecting and monitoring a change by comparing the system dynamics from the identification unit with the past system dynamics recorded in the recording unit, and the change monitoring unit And an abnormality determination unit that determines an abnormality when the change in the system dynamics detected in step 1 deviates from a preset value .
[0013]
Hereinafter, the present invention will be described in detail. In recent years, deterministic chaos has been studied in addition to AI theory and FUZZY theory. This can often produce seemingly irregular, unstable, and complex behaviors from differential and differential equations that are dominated by determinism (if the initial value is determined, all subsequent states are determined in principle) It is derived from the phenomenon.
[0014]
The concept of deterministic chaos revealed that irregular phenomena in the world are not necessarily only those that are dominated by non-determinism. In other words, there are not a few phenomena in the world that produce their behavior according to determinism. Common to dynamical systems that produce chaotic behavior is that their equations are non-linear.
[0015]
By the way, if the behavior of certain time-series data is chaotic, it can be considered that the behavior follows a deterministic law. If it is possible to estimate the nonlinear deterministic law, the near future until the deterministic causality is lost due to the “sensitive dependence on the initial value” of chaos from the measurement data at a certain point in time. It is possible to predict the data.
[0016]
Prediction from the standpoint of such a deterministic dynamical system theory is based on the embedded theory of time-series data that “the state space and attractor of the original dynamical system are reconstructed from a single observation time-series data”. ing. The outline of this theory is as follows.
[0017]
First, as shown in FIG. 4, a vector x (t) shown in the equation (1) is generated from the observed time series data y (t) (τ is a time delay).
[0018]
[Expression 1]
x (t) = (y (t), y (t−τ), y (t−2τ),..., y (t− (n−1) τ))
This vector represents one point in the n-dimensional reconstruction state space R ⁿ . Therefore, as shown in FIG. 5, when t is changed, a trajectory can be drawn in the n-dimensional reconstruction state space.
It is assumed that the target system is a deterministic dynamical system, and the observation time series data is obtained through an observation system corresponding to the C ¹ continuous mapping from the state space of this dynamical system to the one-dimensional Euclidean space R. In this case, this reconstructed trajectory is embedded in the original deterministic system if n is sufficiently large.
[0019]
That is, if the observed time series data is derived from the original dynamical attractor, the attractor that preserves the attractor's phase structure is reproduced in the reconstructed state space. Therefore, x (t) moves on the attractor, and in the short term, a position after s steps of x (t), that is, x (t + s) can be predicted. Here, x (t + s) is represented by the following equation.
[0020]
[Expression 2]
x (t + s) = (y (t + s), y (t + s−τ), y (t + s−2τ),..., y (t + s− (n−1) τ))
Here, since y (t + s), which is a component of x (t + s), is time-series data that is s steps ahead from the observation time point, this value is a predicted value at the s steps ahead. Furthermore, when s> τ, y (t + s−τ) is also a predicted value. Thus, a predicted value ahead of a predetermined step can be obtained from the value of s. However, as described above, the chaotic system is sensitive to the initial value, so it is not suitable for long-term prediction. Conversely, for near-future data until deterministic causality is lost, high-precision prediction is possible.
[0021]
According to the present invention, an attractor is generated in the reconstruction space, short-term prediction is performed, and the presence or absence of an abnormality is determined by comparing the predicted value with the actual measurement value. The presence / absence of abnormality / normality is flexibly determined according to the state.
[0022]
As described above, in the present invention, a predicted value is generated in consideration of the state of the observation system, and an abnormality is determined by comparing the predicted value with the actual measurement value. An abnormal determination is made. Furthermore, since the chaos short-term prediction method is used for the prediction unit, it is possible to obtain a highly accurate prediction for the perturbation of the system.
[0023]
In particular, in the conventional method, when setting a reference value for determining that an observation system is abnormal, the maximum value (or the minimum value) that the observation system can take is used as the reference value, so that it corresponds to the state of the observation system. It was difficult to make an abnormality determination. For example, even when an observation value that is determined to be normal in normal times is obtained, it may be appropriate to determine that an abnormality is present when the state of the observation system is taken into consideration. Thus, in the conventional method, it is difficult to take the state of the observation system into consideration, and even in such a case, it has been determined to be normal.
[0024]
DETAILED DESCRIPTION OF THE INVENTION
A first embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows a functional block diagram of an abnormality monitoring apparatus according to the present invention. In FIG. 1, 1 is a short-term prediction system unit, 2 is a time-series data short-term prediction unit (time-series data storage unit and prediction unit) that also serves as a data file unit, 3 is a data input unit that also serves as a sensor unit, and 7 is a prediction result. An extraction unit, 8 is an abnormality monitoring unit, and 9 is a man-machine unit.
[0025]
The data input from the data input unit 3 is sent to the time-series file 4 and then input to the time-series data short-term prediction unit 2 or the parameter optimization unit 5. When data is input to the parameter optimization unit 5, the parameter optimization unit 5 appropriately optimizes parameters and then inputs the optimized parameter data to the time-series data short-term prediction unit 2.
[0026]
The time-series data short-term prediction unit 2 performs a prediction result n steps ahead based on the input data, parameters, and the like. The prediction value is sent to the prediction value file 6 and then input to the parameter optimization unit 5 and the prediction result extraction unit 7 respectively. The parameter optimization unit 5 optimizes the parameters based on the input time series data in addition to the time series data.
[0027]
The prediction result extraction unit 7 extracts the prediction result from the prediction value file 6 and inputs it to the abnormality monitoring unit 8. The abnormality monitoring unit 8 compares the predicted value with the actual measurement value, determines whether or not an abnormality has occurred based on a predetermined criterion, and displays the result through the man-machine unit 9.
[0028]
Hereinafter, the operation of the short-term prediction system 1 will be described.
[0029]
First, the time series data y (t), y (t−τ), y (t−2τ) (where τ is a time delay) is sent from the data input unit 3 to the time series data short-term prediction unit 2 through the time series file 4. Entered. The time-series data short-term prediction unit 2 reconstructs the n-dimensional reconstructed state space R ⁿ based on the input time-series data, and calculates the predicted value after s steps by reproducing the attractor of the observation system. The predicted value is compared with the actual observed value after the s step. FIG. 2 shows an explanatory diagram thereof.
[0030]
As shown in FIG. 2, when performing abnormality determination, an upper limit value and a lower limit value are set based on a predicted value, and an area between the upper limit value and the lower limit value is set as a normal observation system. The normal area to be determined. Other areas are defined as abnormal areas.
[0031]
For example, a predicted value after s steps is obtained at time ta. This predicted value is indicated by a point a in FIG. 2, and a normal region set from the point a is indicated by a solid line portion at time ta−s in FIG. Since the actual observation value at t−a + s is in the normal region, it is determined that the observation value at this point is in the normal state.
[0032]
Further, the predicted value after step s obtained at time t is indicated by a point b in FIG. 2, and the normal region set from the point b is indicated by a solid line portion at time t + s in FIG. Since the actual observation value at t + s is in the abnormal region, it is determined that the observation value at this point is in an abnormal state.
[0033]
The time series data short-term prediction unit 2 reconstructs the strange attractor by embedding the time series in the normal state, and records the embedding parameters at this time.
[0034]
When system perturbation occurs, the prediction error increases. When the prediction error exceeds a set threshold value, the parameter optimization unit 5 is caused to execute parameter reoptimization. Of the parameters obtained as a result, it can be seen that if the change in “delay” is checked, it becomes “abnormal”. However, the width of the perturbation resulting in “abnormality” / the “delay” of the embedding parameter differs depending on the system to be applied.
[0035]
In the chaos short-term prediction method of this method prediction unit, the time series is embedded and the strange attractor is reconstructed. If you check the shape and density of the attractor, you can understand the state of the system. Taking an electroencephalogram time series as an example, the shape and density of attractors differ depending on the psychological state. Of course, when you are injurious, or when you are mentally ill, the shape is clearly different.
[0036]
Thus, in this embodiment, an appropriate normal area can be set in response to a dynamic change in time series. For example, the normal regions at times ta + s and t + s in FIG. 2 are greatly different, indicating that an optimal normal region can be set according to the state of the observation system.
[0037]
In addition, the normal area | region in a predicted value can also be made variable with the value or the system to apply. For example, take only the lower limit value of the normal region, narrow the specified standard (when the observed value is above a certain value or below a certain value, etc.) narrow the normal region, or normal when the observed value is close to the prescribed value Such as taking a wide area.
[0038]
Thus, the adjustment of the normal region due to the difference in value often follows a rule based on an empirical rule. Fuzzy theory may be applied to such normal / abnormal area allocation. Although the conventional method is also realized by AI / Fuzzy theory, in AI / Fuzzy theory, it is necessary to estimate the abnormal tendency of the state, whereas in this example, the meaning of the Fuzzy rule is clear, Rules become simple.
[0039]
Moreover, the shape of the strange attractor changes depending on the perturbation of the system. Focusing on this point, it is clear that when the strange attractor in the normal state of the system changes, it is moving toward an abnormal state. Therefore, the abnormality determination can be performed by monitoring the state of the attractor.
[0040]
Furthermore, since the chaos short-term prediction method is used for the prediction, it is possible to perform a highly accurate short-term prediction and to obtain a highly accurate prediction corresponding to the perturbation of the system.
[0041]
This abnormality monitoring apparatus can be used for the following applications, for example.
[0042]
(1) Abnormal phenomenon detection device for chaotic time-series data (2) Abnormal vibration detection device for rotating machine (motor, generator, etc.) (3) Congestion detection device for road traffic (4) EEG abnormality detection device ( 5) Turbine shaft vibration abnormality detection device (6) Pulse abnormality detection device (7) Turbine, turbine vibration abnormality detection device FIG. 6 is a functional block diagram of a temperature change pre-detection alarm device showing a second embodiment of the present invention. 6, the data input unit 3 captures data from a target process at a constant time interval (for example, 100 ms, 1 second, 1 minute, etc.). The captured data is added to the time series file 4 as the latest data. Next, a prediction request is issued from the data input unit 3 to the time-series data short-term prediction unit 2. The short-term prediction unit 2 performs a prediction set by the system (for example, temperature prediction one hour ahead) according to a request from the data input unit 3 and stores the result in the prediction value file 6. The predicted value file 6 has the same structure as the time series file 4 ( observed value time series file) and is a time series file of predicted values.
[0043]
Next, the prediction result extraction unit 7 is activated by the time-series data short-term prediction unit 2. The prediction result extraction unit 7 extracts the previous prediction value, that is, the prediction value corresponding to the observation value captured by the data input unit 3 this time from the prediction value file 6 and the observation value from the time series file 4. Then, data that is a set of the predicted value and the observed value is input to the system dynamics identification unit 8a. In the identification unit 8a, the parameter optimization unit 5 identifies the system dynamics based on the attractor by using the parameters that define the system dynamics. Dynamics include the maximum Lyapunov exponent and fractal dimension, prediction results, observations, etc.
[0044]
The identified system dynamics are recorded in the recording file 8d and input to the system dynamics change monitoring unit 8b. The system dynamics change monitoring unit 8b detects past changes in system dynamics. This change detection is performed using absolute value evaluation or% evaluation. The change in system dynamics in the change monitoring unit 8b is given to the abnormality determining unit 8c. If it is determined that the change is abnormal, the change is supplied to the man-machine unit 9 for display processing. Even when there is no abnormality in the change, if the change monitoring unit 8b can determine that there is a change in the dynamics of the process / plant, the parameter optimization unit 5 is started by the change monitoring unit 8b to optimize the parameter. The unit 5 can obtain an optimum parameter expressing the dynamics, and the entire system can perform a monitoring alarm following the change of the plant.
[0045]
According to the 2nd form comprised as mentioned above, the following functions are obtained.
[0046]
(A) automatically identifying system dynamics using only time-series data (optimum parameter estimation);
(B) Make necessary short-term forecasts based on the parameters obtained,
(C) If the prediction result is different from the previous trend, an alarm is issued.
(D) If the prediction result is significantly different from the observed data, the dynamics of the system has changed, so the parameter estimation in (a) is performed again.
[0047]
With the above function, parameter setting for short-term prediction is unnecessary in the second embodiment of the present invention, and it can be used for general purposes.
[0048]
The range of application of the second embodiment extends as follows.
[0049]
(1) Wind direction change pre-detection alarm device (2) Humidity change pre-detection alarm device (3) Flow rate change pre-detection alarm device (4) Electricity change pre-detection alarm device (5) Traffic change pre-detection alarm device (6) Cooling water amount change pre-detection alarm device (7) Gas amount change pre-detection alarm device (8) Other various time series data change detection alarm devices
【The invention's effect】
As described above, in the present invention, the normal range is determined from the predicted value obtained by the chaos short-term prediction method, and by detecting the deviation from the normal range, the abnormality is detected, so that it corresponds to the state of the observation system. It is possible to make an abnormality determination.
[0051]
In particular, in the case of a phenomenon having dynamic dynamic characteristics, there is a possibility that an abnormal state may exist even in a region that has been considered normal, so that it is difficult to detect an abnormality with a method of specifying the region. According to the method, the normal range can be determined corresponding to the change in time series, so that an accurate abnormality determination is possible.
[0052]
Furthermore, if an abnormality is determined, the device can be prevented from being destroyed by performing a system alarm or abnormality processing (stopping, etc.).
[0053]
In addition, there is an advantage that parameter setting for short-term prediction becomes unnecessary.
[Brief description of the drawings]
FIG. 1 is a functional block diagram of an abnormality monitoring apparatus according to a first embodiment of the present invention.
FIG. 2 is a graph showing the relationship between predicted values and observed values.
FIG. 3 is a graph showing a normal region and an abnormal region of observed values.
FIG. 4 is an explanatory diagram of time-series data.
FIG. 5 is an explanatory diagram of an attractor.
FIG. 6 is a functional block diagram of a state change alarm device according to a second embodiment of the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Short-term prediction system 2 ... Time series data short-term prediction part 3 ... Data input part 4 ... Time series file 5 ... Parameter optimization part 6 ... Prediction value file 7 ... Prediction result extraction part 8 ... Abnormality monitoring part 9 ... Man machine part

Claims (1)

A storage unit for time series data of the observation system,
A time vector after a predetermined step is generated by chaos inference based on the behavior of the attractor generated in the reconstruction space by generating a data vector using the time series data obtained from the storage unit as a parameter, and embedding the data vector A prediction unit for obtaining a predicted value of the data;
When the prediction value obtained from this prediction unit is compared with the observation value in the observation system and the prediction error becomes large, a parameter optimization unit that performs parameter optimization processing,
The predicted value and the observed value of the observation system measured after elapse of the predetermined step are extracted and input, and the maximum Lyapunov exponent, fractal dimension, and predicted value based on the attractor using the parameter from the parameter optimization unit And an identification unit for identifying system dynamics consisting of observation values,
A recording unit for recording the system dynamics from the identification unit;
A change monitoring unit that detects and monitors a change in comparison between system dynamics from the identification unit and past system dynamics recorded in the recording unit;
A state change alarm device comprising: an abnormality determination unit that determines an abnormality when a change in system dynamics detected by the change monitoring unit deviates from a preset value.

Images