Anomaly detection method

Abstract

To provide an abnormality detection method that can continue abnormality detection without interfering with principal component analysis even when signal data is deficient.SOLUTION: When signal data required for principal component analysis is deficient, the deficient signal data is estimated and complemented by using normal signal data adjacent to or related to the deficient signal data. When principal component analysis is performed and a principal component vector is calculated, the estimated and complemented signal data is used with intentional variation. When Q statistic is calculated from the principal component vector, the estimated and complemented signal data is used without intentional variation.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an abnormality detection method, particularly an abnormality detection method for detecting an abnormality of equipment from the results of principal component analysis of a plurality of pieces of signal data of the same type in the equipment.

Background Art

[0002] As such an abnormality detection method, for example, there is one described in Patent Document 1 below. This abnormality detection method uses the output signals of a plurality of temperature sensors provided in a mold of a continuous casting facility, and predicts a breakout of the continuous casting facility from the results of principal component analysis of those signal data (temperature data). In this abnormality detection method, a principal component vector is calculated by principal component analysis from the plurality of temperature data of the detected mold, and when the Q statistic using this principal component vector exceeds a predetermined value, it is detected that there is a sign of a breakout and thus an abnormality of the equipment. Since a certain number of data (signal data) is required for principal component analysis, for example, when the signal data is insufficient due to a failure (abnormality) of a temperature sensor or a shortage in the number of temperature sensors due to equipment replacement, a method of estimating and complementing the signal data by interpolation or extrapolation is also described.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the estimated and interpolated signal data may not always be accurate. In such cases, the accuracy of the principal component vector calculation by principal component analysis decreases, the Q statistic becomes larger, and there is a risk of misidentifying equipment abnormalities, i.e., signs of breakout. If signs of breakout are misidentified, the operation of the continuous casting equipment must be stopped, reducing the production efficiency of cast steel. For example, if the lack of signal data is due to a failure of the temperature sensor, the operation of the continuous casting equipment can be avoided by replacing the temperature sensor or the equipment (mold) containing the temperature sensor during the equipment shutdown. However, there are limits to improving the estimation accuracy of the interpolated signal data. In other words, an anomaly detection method is desired that can continue anomaly detection without hindering principal component analysis even when signal data is insufficient.

[0005] The present invention has been made in view of the above problems, and its purpose is to provide an anomaly detection method that can continue anomaly detection without hindering principal component analysis even when signal data is insufficient. [Means for solving the problem]

[0006] To achieve the above objective, an anomaly detection method according to one aspect of the present invention is an anomaly detection method for detecting anomalies in equipment from the principal component analysis results of multiple identical signal data in the equipment, wherein when there is insufficient signal data necessary for the principal component analysis, the insufficient signal data is estimated and supplemented using normal signal data adjacent to or related to the insufficient signal data, and when the principal component analysis is performed and the principal component vector is calculated, the estimated and supplemented signal data is used with intentional variation, and when the Q statistic is calculated, the estimated and supplemented signal data is used without intentional variation.

[0007] Furthermore, a further aspect of the present invention is characterized in that the deficiency of signal data is caused by any one of the following: a sensor failure, a shortage of sensors due to equipment replacement, or a shortage of signal data outputs from sensors due to changes in product dimensions or position. A further aspect of the present invention is characterized in that the signal data is the output signal of a temperature sensor inside the mold of a continuous casting facility, and the abnormality is breakout prediction in the continuous casting facility. [Effects of the Invention]

[0008] According to the anomaly detection method of the present invention, even if signal data is insufficient, principal component analysis is not hindered, and anomaly detection can be continued without reducing the operation of the equipment. [Brief explanation of the drawing]

[0009] [Figure 1] This is a schematic diagram showing one embodiment of a continuous casting facility to which the abnormality detection method of the present invention is applied. [Figure 2] Figure 1 is an explanatory diagram of the mold for a continuous casting machine. [Figure 3] This graph shows an example of temperature data detected by the temperature sensor in Figure 2. [Figure 4] This graph shows another example of temperature data detected by the temperature sensor in Figure 2. [Figure 5] Figures 3 and 4 show graphs of the Q statistic calculated using the temperature data. [Figure 6] This graph shows an example of estimated and interpolated temperature data used to calculate principal component vectors. [Figure 7] This is a graph of the Q statistic calculated using the temperature data in Figure 6. [Figure 8] This graph shows an example of temperature data detected by the temperature sensor in Figure 2. [Figure 9] This is a graph of the Q statistic calculated using the normal temperature data from Figure 8. [Figure 10] Figure 8 shows a graph of the Q statistic calculated by intentionally shifting one of the temperature data points. [Figure 11] This is a graph of the Q statistic calculated by intentionally shifting and introducing variability to one of the temperature data points in Figure 1. [Modes for carrying out the invention]

[0010] The embodiments of the anomaly detection method of the present invention will be described in detail below with reference to the drawings. The embodiments shown below are illustrative examples of devices and methods for realizing the technical concept of the present invention, and the technical concept of the present invention is not limited to the following embodiments in terms of the material, shape, structure, arrangement, etc. of the components. Furthermore, the drawings are schematic. Therefore, it should be noted that the relationship and ratio of thickness and planar dimensions may differ from those in reality, and there may be parts where the dimensional relationships and ratios differ between drawings.

[0011] Figure 1 is a schematic diagram showing one embodiment of a continuous casting equipment to which an abnormality detection method is applied, and is the same as that described in Patent Document 1. This abnormality detection method detects the occurrence of breakout in a continuous casting equipment. In the continuous casting machine 1 of this embodiment, molten steel 2 poured into the tundish 3 is poured into the mold 5 via an immersion nozzle 4, and the semi-solid slab 6 withdrawn from the mold 5 is transported while being supported by a plurality of slab support rolls 7. A plurality (many) temperature sensors 8 are embedded in the mold 5, and based on the temperature of the mold 5 detected by these temperature sensors 8, the determination unit 20 detects signs of breakout.

[0012] Figure 2 shows an example of the arrangement of temperature sensors 8 embedded in the mold 5. The cast slab 6 is withdrawn in the casting direction A, and multiple (many) temperature sensors 8 are arranged at equal intervals in the mold 5 at a certain position, i.e., at a certain height, in this casting direction A. These temperature sensors 8 are embedded at a predetermined depth from the outer wall surface of the mold 5, and in practice, they are arranged in multiple rows in the height direction, for example, three rows. For information on the breakout precursor phenomenon, please refer to Patent Document 1 mentioned above, but in short, while the temperature of the molten steel 2 or cast slab 6 detected at the same height position is almost constant, an imbalance occurs in the temperature detected at the same height position as a precursor to breakout. Therefore, principal component analysis is performed on the temperature (temperature data) detected at the same height position, and the amount of deviation from the principal component vector is calculated as the Q statistic. When this Q statistic exceeds a predetermined threshold, it is detected as a precursor to breakout. Furthermore, since the continuous casting machine 1 also produces slabs 6 with different specifications, such as plate width, in order to continue detecting signs of breakout even for such slabs 6 with different specifications, the temperature data detected by the temperature sensor 8 is interpolated so that the number of data points for principal component analysis (the number of calculation cells) remains constant, and principal component analysis is performed, and the Q statistic is calculated from the principal component vector.

[0013] When calculating the Q statistic from the principal component analysis results, for example, a temperature sensor 8 installed in the mold 5 may malfunction and detect abnormal temperature data. When such abnormal temperature data intervenes in the principal component analysis, it becomes impossible to accurately determine the principal component vectors, and the Q statistic may become unnecessarily large, potentially leading to the false detection of an anomaly, i.e., a breakout precursor. Also, a malfunction of temperature sensor 8 may result in insufficient temperature data. As mentioned above, the same number of temperature data is required for principal component analysis, so in the above-mentioned Patent Document 1, the insufficient temperature data is estimated and supplemented from, for example, the temperature data of an adjacent temperature sensor 8 or a temperature sensor 8 that is considered to have a high correlation, and the principal component analysis continues. If a genuine breakout precursor is detected, it is necessary to reduce the operation of the continuous casting equipment, but if, for example, only a temperature sensor 8 malfunctions, there is a need to estimate and supplement the temperature data of the malfunctioning temperature sensor 8 and continue detecting breakout precursors. The malfunctioning temperature sensor 8 can be replaced, for example, when the equipment is shut down.

[0014] However, the estimated and interpolated temperature data is not necessarily accurate enough to allow for the accurate calculation of principal component vectors using principal component analysis with estimated and interpolated temperature data. As a result, the Q statistic may become large, potentially leading to false detection of breakout precursors. For example, Figure 3 shows the signal data, i.e., temperature data, of a faulty temperature sensor 8, while Figure 4 shows the temperature data of a non-faulty temperature sensor 8 adjacent to the faulty temperature sensor 8 (hereinafter referred to as the adjacent temperature sensor 8). As is clear from the figures, the temperature data in Figure 3 shows a significant fluctuation compared to the temperature data in Figure 4. Normally, the temperature data of a faulty temperature sensor 8 would show a value similar to the temperature data in Figure 4. In this example, the temperature data of the faulty temperature sensor 8 shown in Figure 3 showed a significant fluctuation from, for example, time t0. The Q statistic was calculated from the principal component analysis results using this temperature data up to time t1, and after time t1, the Q statistic was calculated from the principal component analysis results (influence coefficients) prior to time t1. The calculation results of the Q statistic for this time period are shown in Figure 5. As is clear from the figure, the Q statistic is calculated to be large not only when using temperature data with significant fluctuations from the faulty temperature sensor 8, but also when using temperature data (constant value) from the adjacent temperature sensor 8. Normally, the Q statistic at any given time would exceed the breakout indicator threshold.

[0015] In contrast, FIG. 6 shows temperature data with intentional variations given to the temperature data of the adjacent temperature sensor 8 shown in FIG. 4. The temperature data with such variations is obtained by adding approximately the same amplitude at approximately the same short period to the temperature data shown in FIG. 4. In this example, assuming that the failure of the temperature sensor 8 that occurred at time t0 has been detected, principal component analysis is performed using the temperature data in FIG. 6 instead of the temperature data of the failed temperature sensor 8 from this time t0 to obtain a principal component vector, and the deviation value with respect to this principal component vector, that is, the Q statistic, is calculated. The failure of the temperature sensor 8 may be detected, for example, when a temperature region that is impossible if normal is detected, or when a temperature that fluctuates at a physically inconceivable speed (timing, temperature change rate) is detected. Note that when calculating the Q statistic, temperature data with variations removed (= the temperature data in FIG. 4) is used. The calculation result of the Q statistic for this time period is shown in FIG. 7. As is clear from the figure, the Q statistic remains at a small value over all time periods, and therefore does not exceed the breakout prediction threshold, and the breakout prediction detection can be continued. This is because, due to the intentional variations given to the temperature data, the scale of the variable as a result of standardization becomes smaller compared to other variables, and thus the influence coefficient of the corresponding variable of the principal component vector when performing principal component analysis becomes relatively small, and the sensitivity to anomaly detection can be intentionally reduced. In this example, since there is only one piece of missing temperature data, it is sufficient to simply give variations using the above method. However, when there are multiple signal data to be estimated and supplemented, for example, it is desirable to change the period and amplitude of the above method for each data so that there is no correlation between those data.

[0016] Next, we will describe a different example in which principal component analysis is performed on estimated and interpolated temperature data with intentional variability, and the Q statistic is calculated from the resulting principal component vector. Figure 8 shows an example of temperature data actually detected by temperature sensor 8. That is, the temperature data shown in this example is normal temperature data detected by a non-faulty temperature sensor 8. In calculating the Q statistic using principal component analysis, many more temperature data points (50 points as an example) are used, but Figure 9 shows the results of performing principal component analysis using these normal temperature data and calculating the Q statistic from the resulting principal component vector. Because the scale of the vertical axis is small, the Q statistic appears large, but the actual value is small. The following simulation attempts to see the effect of the temperature data shifted from the normal value when one of the temperature data points in Figure 8 (actually one of many temperature data points) is shifted from the normal value (by performing low-precision estimation interpolation), and then the temperature data is returned to the normal value in Figure 8, and the Q statistic is calculated from the previous principal component vector.

[0017] First, assuming that one of the thermometers fails and the estimated accuracy is low, the adjacent temperature data is multiplied by 1.2 and principal component analysis is performed. Then, at time t2, the thermometer is restored to normal, and the calculation results when calculating the Q statistic from the previous principal component vector are shown in Fig. 10. As is clear from Fig. 10, when principal component analysis is performed using the adjacent temperature data of the failed thermometer multiplied by 1.2, and after that temperature data is restored to normal, the Q statistic obtained from the previous principal component vector deviates from the Q statistic based on the normal temperature data in Fig. 9. In contrast, the adjacent temperature data (the same data as above) of the failed thermometer is multiplied by 1.2, and further intentional variations (±10°C) are given to this data and principal component analysis is performed. Then, at time t2, the thermometer is restored to normal, and the calculation results when calculating the Q statistic from the previous principal component vector are shown in Fig. 11. As is clear from the figure, compared with the case where only the adjacent temperature data is multiplied by 1.2, when intentional variations are given to the 1.2-fold adjacent temperature data, the calculated Q statistic hardly changes from the Q statistic obtained using normal temperature data. In this case, if the absolute value of the variation set to ±10°C is increased, the deviation amount from the Q statistic obtained using normal temperature data can be further reduced. On the other hand, if the size of the variation is reduced, the deviation amount from the normal Q statistic can also be increased.

[0018] The above has described the estimation and complementation of temperature data assuming the case where the temperature sensor 8 fails, but it can be widely applied to the case of estimating and complementing the insufficient part when the temperature data required for principal component analysis is insufficient. For example, it can also be used when the number of temperature sensors changes (decreases) due to equipment replacement or the like. Or it is also applicable to the insufficient output number of signal data from temperature sensors due to dimensional variations or position variations of products, that is, slab.

[0019] Although the anomaly detection method according to the embodiment has been described above, the present invention is not limited to the configuration described in the above embodiment, and various modifications are possible within the scope of the gist of the present invention. For example, in the above embodiment, only an anomaly detection method for detecting signs of breakout in a continuous casting facility as an anomaly was described in detail, but the anomaly detection method of the present invention can be applied to any method that performs principal component analysis using multiple (many) identical signal data from the equipment, calculates a Q statistic from the principal component vector, and performs anomaly detection.

[0020] Thus, in this embodiment, if there is insufficient signal data necessary for principal component analysis, the missing signal data is estimated and interpolated using normal signal data adjacent to or related to the missing signal data. When performing principal component analysis and calculating the principal component vectors, the estimated and interpolated signal data is used with intentional variation. When calculating the Q statistic from the principal component vectors, the estimated and interpolated signal data is used without intentional variation. This makes it possible to continue detecting abnormalities in the equipment, for example, even if a sensor that outputs signal data malfunctions and the accuracy of the estimated and interpolated output data from that malfunctioning sensor is not sufficient.

[0021] Furthermore, it can handle signal data shortages caused by sensor failure, a shortage of sensors due to equipment replacement, or a shortage of signal data output from sensors due to changes in product dimensions or position. This expands its applicability to address signal data shortages and improves the continuity of equipment anomaly detection. [Explanation of symbols]

[0022] 1. Continuous casting machine 2 Molten steel 3 Tan Dish 4 Immersion nozzle 5. Mold 6 cast slabs 7. Cast slab support rolls 8. Temperature sensor 20 Judgment section

Claims

1. An anomaly detection method for detecting anomalies in equipment based on the principal component analysis results of multiple identical signal data in the equipment, An anomaly detection method characterized in that, when there is insufficient signal data necessary for the principal component analysis, the missing signal data is estimated and supplemented using normal signal data adjacent to or related to the missing signal data, the principal component analysis is performed and the estimated and supplemented signal data is used with intentional variation, and when calculating the Q statistic, the estimated and supplemented signal data is used without intentional variation, and the intentional variation is a waveform with periodicity and variation with a regularity of similar amplitude.

2. The abnormality detection method according to claim 1, characterized in that the shortage of signal data is caused by one of the following: a sensor failure, a shortage of sensors due to equipment replacement, or a shortage of signal data outputs from sensors due to changes in product dimensions or position.

3. The abnormality detection method according to claim 1 or 2, characterized in that the signal data is the output signal of a temperature sensor inside the mold of a continuous casting facility, and the abnormality is breakout prediction in the continuous casting facility.

Images