JP2022050354A - Feature quantity selection method and health state evaluation method - Google Patents

Abstract

To heighten the high efficiency, reliability and universality of a feature quantity selection method.SOLUTION: The present invention pertains to a feature quantity selection method and a health state evaluation method. The feature quantity evaluation method includes: a data acquisition step for acquiring a plurality of feature values of a measurement object, for each of a plurality of feature quantities, the feature value representing a feature quantity; a feature quantity evaluation step for calculating, for each of feature quantities waiting for evaluation among the plurality of feature quantities, the non-monotonic index and deviation index of the feature quantity on the basis of the plurality of feature values obtained in the data acquisition step, the non-monotonic index reflecting the non-monotonicity of feature value, the deviation index reflecting the degree of divergence of feature value with respect to the change tendency of the whole; and a feature quantity selection step for calculating, for each of the feature values waiting for evaluation, an overall index on the basis of the non-monotonic index and the deviation index, and selecting the optimum feature quantity suitable for the evaluation of the health state of the measurement object on the basis of a preliminarily set index threshold.SELECTED DRAWING: Figure 7

Description

The present invention relates to a feature quantity selection method for selecting a feature quantity suitable for evaluating the health condition of a subject to be measured from a plurality of feature quantities and a health condition evaluation method for evaluating the health condition of the subject to be measured.

With the development of society, more and more equipment and tools are entering people's lives and production activities, and play an increasingly important role in them. In order to guarantee the moderateness of production and living facilities, evaluation of the health condition of equipment, equipment parts, tools, etc. is an urgent issue. Faced with different equipment, parts and tools, people have submitted a variety of different ways to assess their health. For the health status of many subjects (eg, remaining life), the features that appear cannot or are difficult to measure directly, so the most commonly used method today is to use sensors to provide indirect signals of the condition of the subject under test during its operating process. It is acquired and then analyzed for the health condition of the subject to be measured based on the signal data.

For example, Patent Document 1 discloses a method and system for predicting the life of a blower component, and when predicting the life of a blower component, first collects characteristic parameter information of the blower component, and the characteristic parameter is mainly an electrical characteristic parameter ( For example, it was pointed out that it includes current, voltage, frequency, etc.), mechanical feature parameters (eg, vibration signal, magnetic field signal, temperature signal, etc.), and process parameters (eg, wind speed, work rate, temperature, etc.). It was pointed out that at least one of these sensors would be used in various uses, and then the support vector method would be used to predict the remaining life of the component.

Chinese Patent Application Publication No. 103019135

In the technical proposal for evaluating the health condition of the object to be measured as described above, the selection and mounting method of the sensor is usually determined by the type of the object to be measured and the type of operation method, but an indirect monitoring / measurement method. However, the signal output from the sensor often has interference information, and the amount of data is enormous. If the sensor signal is directly input to the learning algorithm, the analysis result will be affected by the interference information and the accuracy will be improved. It is not expensive and the calculation efficiency is not good.

On the other hand, the technical proposal of Patent Document 1 does not point out the format and specific contents of the input information of the support vector method. In practice, selecting a sensor signal and extracting useful information from the signal is the key to guaranteeing prediction accuracy. The signal collected in Patent Document 1 often contains various types of interference noise, and for example, the vibration signal inevitably includes external vibration interference. If these signals are directly input to a learning algorithm such as a support vector as input data, the prediction result of learning will be interfered with by noise and the accuracy will be reduced.
Therefore, in many cases, in order to guarantee learning accuracy and calculation speed, after acquiring the sensor signal, an expert analyzes the signal data, and further, the health condition of the object to be measured is accurately determined from the signal data. It is necessary to extract and select the feature amount suitable for evaluation. After selecting the features, the health condition of the object to be measured is evaluated by inputting the feature values corresponding to these features into the learning algorithm.

From this, it can be seen that whether or not the selection of features is appropriate directly affects the accuracy of the final evaluation result.

However, in the above analysis process, it is necessary for the expert not only to be familiar with the process of changing the health condition of the object to be measured, but also to have the specialized knowledge of signal analysis. Moreover, even if the objects to be measured are of the same type, the feature amount to be selected may differ depending on the operating conditions.

Therefore, in such an evaluation method, in order to guarantee the accuracy of selection for the feature amount, there is a high demand for the individual ability of the expert, and the selection of the feature amount requires a large amount of man-hours depending on the object to be measured and the operating conditions. Need to be consumed. In addition, the non-uniformity of the level of each expert makes the selection result of the feature amount different, and the selection result is inevitably influenced by subjective factors. Therefore, the conventional technical proposal is expensive, and is not efficient, reliable, and universal.

The present invention has been made to solve the above technical problems, and is used in a feature quantity selection method for selecting a feature quantity suitable for evaluation of a health condition of a subject to be measured from a plurality of feature quantities. For each of the data acquisition step in which a plurality of feature values of the object to be measured are acquired and the feature values are numerical values for expressing the feature amount, and the feature amount awaiting evaluation of each of the plurality of feature amounts. , The non-monotonicity index and the deviation index of the feature amount are calculated based on the plurality of feature values acquired in the data acquisition step, and the non-monotonicity index reflects the non-monotonicity of the feature value. The deviation index calculates a comprehensive index based on the non-monotonic index and the deviation index for each of the feature amount evaluation step that reflects the degree of deviation of the feature value with respect to the overall change tendency and the feature amount waiting for evaluation. Then, a feature amount selection method characterized by including a feature amount selection step for selecting an optimum feature amount suitable for evaluation of the health condition of the object to be measured based on a preset index threshold value is provided. do.

As a result, for each feature amount, the correlation between the feature amount and the health condition of the object to be measured is analyzed based on the collected feature value and the health condition value, and the measurement is performed based on a preset threshold value. It is possible to select a feature amount suitable for evaluation of the health condition of the subject. By defining an index to describe the correlation between the feature amount and the health condition, the optimum feature amount is automatically calculated by designing the quantification calculation and evaluation method for the superiority or inferiority of each feature amount. Not only can the labor consumption be reduced by selecting and artificially analyzing the superiority or inferiority of the feature amount, but also the problem that the selection quality of the feature amount varies due to non-uniformity of the expert level, subjective factors, etc. can be avoided. The high efficiency, reliability and universality of the feature quantity selection method have been improved.

In the feature amount selection method, preferably, in the data acquisition step, a plurality of feature values of one object to be measured are acquired in chronological order for each of the plurality of feature amounts, and in the feature amount evaluation step, a plurality of feature values are acquired. Based on a plurality of feature values acquired in chronological order for one object to be measured in the data acquisition step, the non-monotonicity index and the deviation index are calculated as the first-class non-monotonicity index and the first-class deviation index, respectively. , The comprehensive index is calculated based on the first-class non-monotonic index and the first-class deviation index, and the index threshold is the threshold set for the comprehensive index and the first-class non-monotonic index. Includes at least one of the threshold set in the above and the threshold set for the first-class deviation index.

As a result, when it is difficult to carry out the second-class evaluation, which will be described later, the feature amount can be selected by using the first-class evaluation independently. That is, when it is difficult to acquire the health condition value of the object to be measured, it is possible to select a suitable feature amount based only on the feature value, and the difficulty of data acquisition is reduced.

In the feature amount selection method, preferably, in the data acquisition step, a plurality of feature values of one object to be measured and a corresponding health condition value are acquired for each of the plurality of feature amounts, and the feature amount is described. In the evaluation step, the non-monotonicity index and the deviation index are set as the first-class non-monotonicity index and the corresponding health condition value, respectively, based on the plurality of feature values and the corresponding health condition values acquired for one object to be measured in the data acquisition step. Calculated as a first-class deviation index, the comprehensive index is calculated based on the first-class non-monotonic index and the first-class deviation index, and the index threshold is a threshold set for the comprehensive index, the first. It includes at least one of a threshold set for a first-class non-monotonic index and a threshold set for the first-class deviation index.

As a result, it is possible to deal with a case where the number of objects to be measured is small, and it is sufficient to acquire a plurality of feature values and corresponding health condition values from a single object to be measured for each feature amount.

In the feature amount selection method, preferably, in the data acquisition step, one or a plurality of feature values and corresponding health for each of the plurality of objects to be measured of the same type for each of the plurality of feature amounts. The state value is acquired, and in the feature amount evaluation step, one or more feature values acquired for each of the plurality of objects to be measured of the same type in the data acquisition step and the corresponding health state value are not used. The monotonicity index and the deviation index are calculated as the second-class non-monotonicity index and the second-class deviation index, respectively, and the comprehensive index is calculated based on the second-class non-monotonicity index and the second-class deviation index. The index threshold includes at least one of a threshold set for the comprehensive index, a threshold set for the second-class non-monotonic index, and a threshold set for the second-class deviation index.

As a result, the robustness of the sorting method can be enhanced, the adverse influence on the sorting due to accidental factors is reduced, and the sorting result becomes more reliable.

In the feature amount selection method, preferably, in the data acquisition step, one or a plurality of feature values and corresponding health for each of the plurality of objects to be measured of the same type for each of the plurality of feature amounts. The state value is further acquired, and in the feature amount evaluation step, based on one or more feature values acquired for each of the plurality of objects to be measured of the same type and the corresponding health state value in the data acquisition step. Further, the non-monotonicity index and the deviation index are calculated as the second-class non-monotonicity index and the second-class deviation index, respectively, and the comprehensive index is the first-class non-monotonicity index, the first-class deviation index, and the second-class deviation index. Calculated based on the class 1 non-monotonic index and the class 2 deviation index, the index threshold is a threshold set for the comprehensive index, a threshold set for the class 1 non-monotonic index, and the first. It includes at least one of a threshold set for the class 2 deviation index, a threshold set for the class 2 non-monotonic index, and a threshold set for the class 2 deviation index.

As a result, both the first-class index and the second-class index are taken into consideration, the correlation between the feature amount of the object to be measured and the health condition and the robustness are compatible, and the reliability of the selection result becomes higher.

ここで、ｎはデータ点の数、ｙｉは前記中心線上のｉ番目のデータ点の特徴値、Ｄｉは前記中心線上のｉ番目のデータ点と前記上包絡線又は前記下包絡線上の同一の横座標のデータ点との距離である。 In the feature amount selection method, preferably, in the feature amount evaluation step, the non-monotonic index and the deviation index of the feature amount are calculated according to the following steps for each feature amount waiting to be evaluated.
Step 1. A coordinate system in which the feature value of the feature amount acquired in the data acquisition step, or the feature value of the feature amount and the corresponding health state value, have the feature amount as the vertical axis and the health state or the number as the horizontal axis. Plot with to make the original data, and the numbers represent the time order to acquire the feature values.
Step 2, draw the upper and lower envelopes of the original data,
Step 3, calculate the center line based on the upper envelope and the lower envelope.
Step 4, Based on the center line, the non-monotonic index is calculated according to the following equation (2), and the deviation index is calculated according to the following equation (3).

Here, n is the number of data points, y is the feature value of the i-th data point on the center line, and Di is the same lateral line as the i-th data point on the center line and the upper envelope or the lower envelope. It is the distance from the data point of the coordinate.

Thereby, by defining the calculation method of the non-monotonic index and the deviation index, it is possible to indicate whether or not the feature quantity is suitable for the evaluation of the health condition of the object to be measured in the form of quantification.

In the feature quantity selection method, preferably, in step 2, the upper envelope and the lower envelope of the original data are drawn according to the following five rules.
Rule 1: The first point and the last point of the original data belong to the upper envelope and the lower envelope at the same time.
Rule 2: For other data points other than the first and last points of the original data, if the quadratic derivative with respect to the abscissa is zero, the data points are simultaneously the upper envelope and the lower. Belongs to the envelope,
Rule 3: For the second data point from the back, if the quadratic derivative for its abscissa is greater than zero, then the data point belongs to the lower envelope and if the quadratic derivative for its abscissa is less than zero. For example, the data point belongs to the upper envelope,
Rule 4: For the first point, the second from the back, and the other data points except the last point, if the quadratic derivative for the horizontal coordinate is not zero, then the quadratic derivative of the data point Calculate the product of the next data point with the quadratic derivative, and if the product is greater than zero, then the data point belongs to the upper and lower envelopes at the same time, and if the product is less than zero. If the quadratic derivative of the data point is greater than zero, then the data point belongs to the lower envelope, and if the product is less than zero and the quadratic derivative of the data point is less than zero, then the data. The point belongs to the upper envelope and
Rule 5: When the data points of the original data corresponding to some horizontal coordinates are assigned only to the upper envelope or the lower envelope and the data points of the other envelope are lost, before and after the data points. The data points of the envelope are complemented by performing linear interpolation using the numerical values of the points adjacent to.

Thereby, by using the above-mentioned envelope drawing method, it is possible to reduce the influence on the whole of the local minute change in the data line, and it is possible to surely reflect the whole change tendency.

In the feature amount selection method, preferably, after the data acquisition step, the feature amount simplification step before the feature amount evaluation step is further included, and in this feature amount simplification step, the plurality of feature amounts are evaluated. As a waiting feature amount, a linear correlation coefficient between these two feature amounts is calculated for any two feature amounts in the two feature amounts based on a plurality of feature values acquired in the data acquisition step and corresponding health condition values. When the calculated and calculated linear correlation coefficient exceeds the preset correlation threshold, it is determined that the two features are substantially the same, and the two features are selected from the features waiting to be evaluated. Of these, the feature amount with a relatively long calculation time of the feature value is deleted. As a result, by deleting unnecessary duplicate features, the uniqueness of the features is guaranteed and the amount of calculation is reduced.

In the feature amount selection method, preferably, in the data acquisition step, a sensor signal value is acquired from each sensor for detecting the feature amount installed on the object to be measured, and the acquired sensor signal value is obtained. The preprocessing of the above is performed, and then the feature value is extracted from the sensor signal value after the preprocessing.

As a result, the problems of reduction of accuracy and reduction of calculation efficiency are avoided by reducing the influence of defects on the feature quantity selection of interference information such as noise via preprocessing.

In the feature quantity selection method, the health condition is preferably any one of cumulative operating time, remaining life, and wear amount.

In the present invention, in the health condition evaluation method for evaluating the health condition of the object to be measured, each step of the above-mentioned feature amount selection method and a part or all of the feature amount from the optimum feature amount are used as the feature amount for model construction. A model that selects and builds a mathematical model between the model-building features and the health status based on the feature values and the corresponding health status values acquired for each model-building feature in the data acquisition step. Health condition evaluation to evaluate the health condition of the object to be measured based on the construction step, the feature value of the feature quantity for model construction detected by detecting the object to be measured, and the mathematical model constructed in the model construction step. Further provided are health assessment methods characterized by including, and.

As a result, the feature quantity suitable for evaluation to the health condition of the object to be measured is automatically extracted from the input signal, and the mathematical model is constructed using the historical data to determine the health condition of the object to be measured. It can be evaluated, thereby reducing the number of requests for analysis by experts and increasing the reliability and universality of the evaluation. Therefore, based on the evaluation results, it is possible to monitor and measure the health condition of equipment, parts, etc., and perform accurate maintenance.

In the health condition evaluation method, preferably, in the model construction step, by sorting the optimum feature amount according to the superiority or inferiority of the comprehensive index, the feature amount having a predetermined ratio in which the comprehensive index is relatively good is selected. This is the feature quantity for model construction.

This enhances the rationality and flexibility of constructing the above mathematical model.

FIG. 1 is a flowchart of a feature amount selection method according to the first embodiment of the present invention. FIG. 2 is a diagram showing a method of drawing an upper envelope and a lower envelope. FIG. 3 is a diagram showing a method of calculating the center line based on the upper envelope and the lower envelope. FIG. 4 is a diagram showing a drawn center line. FIG. 5 shows a diagram for calculating a non-monotonic index and a deviation index based on the center line. FIG. 6 is a flowchart of a feature amount selection method according to a modified example of the first embodiment of the present invention. FIG. 7 is a flowchart of the health condition evaluation method according to the fifth embodiment of the present invention. FIG. 8 is a hardware diagram showing a configuration example of a computer that executes each process according to the embodiment of the present invention.

[First Embodiment]
This embodiment relates to a feature amount selection method for selecting a feature amount suitable for evaluation of the health condition of the object to be measured from a plurality of feature amounts, using an evaluation of the remaining life of a tool of an NC machine tool as an example, and a current sensor. By monitoring and measuring the operating state of the tool with sensors such as vibration sensor, acoustic emission sensor, and force sensor, the wear condition of the tool is evaluated, and the time to replace the tool is determined.

At this time, the "measured object" is a tool, and the amount of wear of the tool or the remaining life of the tool can be used to indicate the health condition of the tool. In this embodiment, the amount of wear of the tool measured by a microscope is used. The "health condition" is defined, and the feature quantity extracted from the detected values of the above sensors is defined as the "feature quantity" for evaluating the health condition of the object to be measured.
The feature quantities include the mean value, maximum value, squared mean square root value, amplitude value, peak value, dispersion value, working frequency amplitude value, double working frequency amplitude value, sharpness coefficient, peak factor, and standard of the original signal of the sensor. It may be a feature amount such as a deviation coefficient, or may be the above-mentioned various feature amounts of a new signal after undergoing secondary data processing such as an absolute value of an original signal and a wrapping signal.

FIG. 1 is a flowchart of a feature amount selection method according to the present embodiment.

First, in step S101, a data acquisition step is executed, and a plurality of feature values of the tool, that is, numerical values for expressing the feature quantities are acquired for each of the plurality of feature quantities.

Specifically, for each of the plurality of feature quantities, the plurality of feature values of one tool and the corresponding wear amount are acquired. The feature value and the amount of wear may include data of the full life cycle of the tool (that is, from zero wear to complete wear), or may include only a part of the data in the life cycle.

Next, step S102 is entered, the feature quantity evaluation step is executed, and the plurality of feature values and correspondences acquired for one tool in step S101 for each feature quantity awaiting evaluation among the plurality of feature quantities. Based on the amount of wear, an index showing the correlation between the feature amount and the wear amount of the tool is calculated, and such an index includes a non-monotonic index and a deviation index, in which the non-monotonic index is a feature value. Reflecting the non-monotonicity of, the deviation index reflects the degree of deviation of the feature value from the overall tendency of change. In this embodiment, in order to study the correlation between each feature amount and the tool wear amount for a single tool, the calculated non-monotonicity index and deviation index are referred to as "Class 1 non-monotonicity index" and respectively. It is called "Class 1 deviation index", and "Class 1" here represents the basic evaluation.

ここで、ｎはデータ点の数、ｙｉは中心線上のｉ番目のデータ点の特徴値、Ｄｉは中心線上のｉ番目のデータ点と上包絡線又は下包絡線上の同一の横座標のデータ点との距離である。 Hereinafter, the non-monotonic index and the deviation index are defined, and the calculation method of the non-monotonic index and the deviation index is divided into the following four steps.
For each feature amount awaiting evaluation in step 1, as shown in FIG. 2, the feature value and the corresponding wear amount of the feature amount acquired in step S101 are set on the vertical axis and the wear amount is on the horizontal axis. Plot the data points P1, P2, P3 ... P7 (in the figure, the case of 7 data points is shown, but the number of data points is not limited to this), and a polygonal line is obtained. Connect with and use as original data.
Step 2. Draw the upper and lower envelopes of the original data, and the specific method of drawing the upper and lower envelopes will be described later.
As shown in step 3 and FIG. 3, the center line is calculated based on the upper envelope and the lower envelope, and there may be various methods for calculating each data point of the center line. For example, the calculation method of the following equation (1) may be used.

The ordinate value of the data point on the center line = (the ordinate value of the upper envelope data point in the same abscissa + the ordinate value of the lower envelope data point in the same abscissa) / 2
Equation (1)

As a result, the obtained center line as shown in FIG. 4 is used as a correlation line between the feature amount and the wear amount of the tool.

As shown in Step 4 and FIG. 5, the non-monotonic index is calculated according to the following equation (2) based on the center line, and the deviation index is calculated according to the following equation (3).

Here, n is the number of data points, y is the feature value of the i-th data point on the center line, and Di is the i-th data point on the center line and the data point having the same horizontal coordinates on the upper envelope or the lower envelope. Is the distance to.

For non-monotonic indicators, if the centerline is completely monotonic, the calculated non-monotonic indicator is zero. If it is not completely monotonic, the calculated non-monotonicity index is greater than zero, and the larger the value, the worse the monotonicity of the centerline, and the monotonicity of the feature and the tool wear. Correlation becomes worse. If the calculated non-monotonic index is 1, it means that it is completely non-monotonic.

With respect to the deviation index, in an ideal state, the calculated deviation index is zero, that is, the feature amount can accurately describe the wear amount of the tool, and there is no deviation. The larger the deviation index, the greater the variation in the overall change tendency of the feature value of the feature amount, that is, the feature value becomes unstable, and the accuracy of the wear amount of the tool described by the feature amount deteriorates. Is shown. If the deviation index is equal to 1, it means that the centerline fitting is meaningless because the variability of the data is so great.

The first-class non-monotonic index and the first-class deviation index have been defined above, but the specific calculation method is not limited to the above steps 1 to 4, and the first-class non-monotonic index is next to the feature value. The calculation method is not limited to this as long as it reflects the non-monotonicity with respect to the axis. Similarly, the calculation method is not limited to this as long as the first-class deviation index reflects the degree of deviation of the feature value from its overall tendency of change. Not limited to.

Further, regarding the upper envelope and the lower envelope, in step 2 above, as shown in FIG. 2, the upper envelope and the lower envelope of the original data can be drawn according to the following five rules.
Rule 1: The first point (for example, P1) and the last point (for example, P7) of the original data belong to the upper envelope and the lower envelope at the same time.
Rule 2: For other data points except the first and last points of the original data, if the quadratic derivative with respect to the abscissa is zero, the data point (eg, P2) is It belongs to the upper envelope and the lower envelope at the same time.
Rule 3: For the second data point from the back, if the quadratic derivative for its abscissa is greater than zero, then the data point belongs to the lower envelope and is quadratic for its abscissa. If the derivative is less than zero, then the data point (eg, P6) belongs to the upper envelope.
Rule 4: For the first point, the second from the back, and the other data points except the last point, if the quadratic derivative with respect to the horizontal coordinate is not zero, the quadratic of the data point. Calculate the product of the derivative and the quadratic derivative of the next data point, and if the product is greater than zero, then the data point (eg, P3) belongs to the upper and lower envelopes at the same time, and if. If the product is less than zero and the quadratic derivative of the data point is greater than zero, then the data point belongs to the lower envelope (eg, P5) and if the product is less than zero and the quadratic of the data point is quadratic. If the derivative is less than zero, then the data point (eg, P4) belongs to the upper envelope.
Rule 5: When the data points of the original data corresponding to some horizontal coordinates are assigned only to the upper envelope or the lower envelope and the data points of the other envelope are lost, they are adjacent to each other before and after the data points. The data points of the envelope are complemented by performing linear interpolation using the numerical values of the points (for example, P4 _bottom , P5 _upper , P6 _bottom ).

The above is the definition of how to draw the upper and lower envelopes of the original data, but the specific drawing method is not limited to the above five rules, but should be defined using various conventional methods of drawing envelopes. Can be done.

Finally, step S103 is entered, a feature amount selection step is executed, a comprehensive index is calculated for each of the feature amounts awaiting evaluation based on the above-mentioned non-monotonic index and deviation index, and a preset index threshold value is calculated. The optimum feature amount suitable for evaluation of the wear amount of the tool is selected based on the above.
In this embodiment, what is studied is the correlation between each feature amount and the wear amount of the tool. Therefore, the comprehensive index is calculated based on the first-class non-monotonicity index and the first-class deviation index, and is concrete. The calculation method is not limited, and for example, when the above definition is used, it is optimal when the index value is zero, and it is shown that the larger the numerical value, the worse the correlation. Therefore, the comprehensive index may be the sum of the numerical values of the first-class non-monotonic index and the first-class deviation index, or may be the sum of the weighted numerical values.

As described above, the index threshold for selecting the set feature amount is among the threshold set for the comprehensive index, the threshold set for the first-class non-monotonic index, and the threshold set for the first-class deviation index. In order to select features that meet the requirements more efficiently and accurately, including at least one of the above, for each of the three indicators, the comprehensive index, the first-class non-monotonic index, and the first-class deviation index. It is preferable to set a threshold value, remove the feature amount whose index value exceeds any one index threshold value, and set the surplus feature amount as the optimum feature amount. As a result, features that are too inferior can be removed, and the efficiency of calculation is improved. Preferably, the index threshold value is 0.1, and the threshold value can be appropriately lowered or raised depending on the actual situation.

In addition, in the above embodiment, in order to evaluate the remaining life of the tool, the amount of wear of the tool is defined as "health state", but the present invention is not limited to this, and for example, the cumulative operating time and remaining life of the tool are defined as "health". It may be "state".

Thereby, according to the present embodiment, for each feature amount, the correlation between the feature amount and the wear amount of the tool is analyzed in advance based on the feature value and the wear amount collected from one tool. Based on the set threshold value, the optimum feature amount suitable for evaluating the wear amount of the tool can be selected. In actual machining, by monitoring and measuring these optimum feature quantities, the amount of wear of the corresponding tool can be known, and the tool can be replaced at an appropriate time.

Specifically, by defining an index for describing the correlation between the feature amount and the wear amount of the tool, by designing a quantification calculation and evaluation method for the superiority or inferiority of each feature amount, it is automatically performed. Not only can the optimum feature amount be selected and the labor consumption due to artificial analysis of the feature amount superiority or inferiority can be reduced, but there is also the problem that the selection quality of the feature amount varies depending on the unevenness of the expert level and subjective factors. It can be avoided and the high efficiency, reliability and universality of the feature quantity selection method are enhanced.

[Modified example of the first embodiment]
The specific configuration of this embodiment is not limited to the above description, and may be other embodiments.

For example, in consideration of the fact that interference information such as noise is often attached to the sensor signal value acquired from each sensor for detecting the feature amount installed on the object to be measured such as a tool, in each of the above steps S101. Preprocessing can be performed on the sensor signal value acquired from the sensor, and then the feature value can be extracted from the sensor signal value after the preprocessing. Such pretreatment may include general processing methods such as zero point shift compensation, high frequency signal or low frequency signal filtration removal.

On the other hand, as shown in FIG. 6, a step S104, that is, a feature amount simplification step may be further included after the step S101 and before the step S102. First, the above-mentioned plurality of features are defined as "features awaiting evaluation", and for any two of them, these two features are based on the plurality of feature values acquired in step S101 and the corresponding wear amount. The linear correlation coefficient between the quantities is calculated, and various conventional mathematical methods may be used to calculate the linear correlation coefficient. When the calculated linear correlation coefficient exceeds the preset correlation threshold value, it is determined that the two features are substantially the same, and the two features are selected from the features waiting to be evaluated. Delete features that have a long feature value calculation time.

This guarantees the uniqueness of the feature amount and reduces the calculation amount. Among them, the correlation threshold value can be set according to the actual need, and may be set to 0.99, for example.

[Second embodiment]
The feature amount selection method according to the present embodiment is different from the first embodiment in the following points.

According to the first embodiment, in step S101 shown in FIG. 2, a plurality of feature values of one tool and a corresponding wear amount are acquired for each of the plurality of feature quantities.

However, in actual operation, for example, when collecting the amount of wear of a tool, every time the numerical value of the amount of wear is collected, the operating tool has to stop the operation, and other equipment is removed or cut. It is difficult or complicated to collect the amount of wear of the tool, that is, the health condition value, because it may be accompanied by the movement of the work.

In response to such technical problems, in the data acquisition step S101 shown in FIG. 2, by the feature amount selection method of the present embodiment, one tool (measurement target) is used in chronological order for each of the plurality of feature amounts. It is only necessary to acquire the feature value of, and it is not necessary to acquire the wear amount (health condition value) of the tool. That is, it is only necessary to arrange the acquired feature values in chronological order.

In this case, in the feature amount evaluation step S102, the non-monotonicity index and the deviation index are based on the same calculation method as in the first embodiment based on the plurality of feature values acquired in chronological order for one tool in step S101. Are calculated as the first-class non-monotonicity index and the first-class deviation index, respectively. At this time, the abscissa in FIGS. 2 to 5 is a number representing the time order of the feature amount instead of the wear amount of the tool. You can adopt it. The time interval for collecting each feature value may be uniform or non-uniform.

As a result, according to the present embodiment, when it is difficult to acquire the health condition value of the object to be measured, an appropriate feature amount can be selected only by the feature value, and the difficulty of data acquisition is reduced.

Since the other configurations of the present embodiment are the same as those of the first embodiment, the related description will be omitted.

[Third Embodiment]
The feature amount selection method according to the present embodiment is different from the first embodiment in the following points.

According to the first embodiment and the second embodiment, in step S101 shown in FIG. 2, data of one tool is acquired for each of a plurality of feature quantities, and the feature quantity selected by this is the health state of the tool. Although it is a feature quantity suitable for evaluation, such a sorting method is not robust because the single tool may be different from other tools due to accidental factors or the like.

For such a technical problem, the index is calculated by fusing the feature values of two or more tools according to the common health condition, that is, the wear amount, by the feature quantity selection method of the present embodiment. Specifically, in the data acquisition step S101 shown in FIG. 2, for each of the plurality of feature quantities, a plurality of feature values and corresponding wear amounts (health state) for each of the plurality of tools (measured targets) of the same type. Value). That is, the difference from the second embodiment is that in the present embodiment, the corresponding wear amount is detected so that the data processing can be performed by sorting and fusing the feature values of different tools according to the size of the wear amount. That's what you have to do. Further, only one feature value and the corresponding amount of wear may be acquired for each tool.

In this case, in the feature amount evaluation step S102, the non-monotonicity index and the deviation index are set based on the one or more feature values acquired for each of the plurality of tools of the same type in step S101 and the corresponding wear amount. calculate.

In the present embodiment, since the calculated index can reflect the correlation and robustness between each feature amount of a plurality of tools and the wear amount, the calculated non-monotonic index and deviation index are referred to as “Class 2”, respectively. They are called "non-monotonic index" and "class 2 deviation index". The "class 2" here is an evaluation performed on the "class 1" in consideration of robustness.

In this case, the above-mentioned comprehensive index can be calculated based on the second-class non-monotonic index and the second-class deviation index, and the index threshold is the threshold set for the comprehensive index and the second-class non-monotonic index. It may contain at least one of the threshold value set for the monotonicity index and the threshold value set for the second type deviation index.

As a result, a reasonably appropriate feature amount can be selected when the correlation and robustness between the feature amount of the object to be measured and the health condition are compatible, and the reliability of the selection result becomes higher.
Since the other configurations of the present embodiment are the same as those of the first embodiment, the related description will be omitted.

[Fourth Embodiment]
The present embodiment further has the following configurations on the first and second embodiments.

That is, in step S101 shown in FIG. 2, for each of the plurality of feature quantities, one or a plurality of feature values and the corresponding wear amount are applied to each of the plurality of tools of the same type as in the third embodiment. Further, in step S102, the non-monotonicity index and the deviation index are second, respectively, based on the one or more feature values and the corresponding wear amounts acquired for each of the plurality of tools of the same type in step S101. Calculated as a class non-monotonic index and a class 2 deviation index.

For a tool, if each feature value acquired for a certain feature amount never acquired the corresponding wear amount at the same time, only the feature value obtained for the corresponding wear amount is the feature value of the other tool. The second-class non-monotonic index and the second-class deviation index can be calculated by fusing, and the surplus feature value of the tool is the first-class non-monotonic index and the first of the feature amount of the tool alone. It can be used to calculate the kind deviation index.

Since the first-class non-monotonic index, the first-class deviation index, the second-class non-monotonic index, and the second-class deviation index were calculated respectively, the comprehensive index is the first-class non-monotonic index, the first-class deviation index, and the first-class deviation index. It can be calculated based on the second-class non-monotonic index and the second-class deviation index, and the specific calculation method is not limited. For example, it may be the sum of these numerical values, and the weighted numerical values may be calculated. It may be Japanese.

The index threshold for selecting the feature amount set as described above is set for the threshold set for the comprehensive index, the threshold set for the first-class non-monotonic index, and the first-class deviation index. It includes at least one of a threshold value, a threshold value set for a second-class non-monotonic index, and a threshold value set for a second-class deviation index. In order to select features that meet the requirements more efficiently and accurately, thresholds are set for each of these five, and features whose index values exceed any one of the index thresholds are removed. , It is preferable to set the surplus feature amount as the optimum feature amount. Preferably, the index threshold value is 0.1, and the threshold value can be appropriately lowered or raised depending on the actual situation.

As a result, a reasonably appropriate feature amount can be selected when the correlation and robustness between the feature amount of the object to be measured and the health condition are compatible, and the reliability of the selection result becomes higher.

[Fifth Embodiment]
The present embodiment relates to a health condition evaluation method for evaluating the health condition of a subject to be measured, and a flowchart thereof is shown in FIG.

Steps S701 to S704 in FIG. 7 are the same as S101 to S104 in the feature quantity selection method of the first embodiment, the second embodiment, or the third embodiment, and the description thereof will be omitted here.

In step S705, the model construction step is executed, a part or all of the feature quantities are selected as the model construction feature quantities from the optimum feature quantities selected in the above feature quantity selection step, and the model construction is performed in the data acquisition step. A mathematical model is constructed between the feature amount for model construction and the wear amount based on the feature value acquired for each feature amount and the corresponding wear amount.

Here, as a method for constructing a mathematical model, a general learning algorithm such as a least squares method, a polynomial fitting, a support vector machine, or an MT (Mahalanobis Taguchi) method may be used.

Next, in step S706, if the mathematical model has already been constructed, the health condition evaluation step is executed for the tool under the same operating conditions. The health state of the tool is based on the detected feature value of the model-building feature amount and the mathematical model built in step S705, with the model-building feature amount detected as the detection-waiting feature amount. Is evaluated, that is, the current amount of wear is calculated.

As a result, according to the present embodiment, it is possible to construct an evaluation model of the health condition based on the selected appropriate feature amount, and it is possible to monitor and measure the health condition of equipment, parts, etc. based on the evaluation result. In addition, we will carry out accurate maintenance.

[Varied example of the fifth embodiment]
The specific configuration of this embodiment is not limited to the above description, and other embodiments may be used.

For example, in the above health condition evaluation method, step S702 is not essential and may be omitted.

Further, in step S705, the model is not constructed using all the optimum feature amounts, but by sorting the optimum feature amounts according to the superiority or inferiority of the above-mentioned comprehensive index, the comprehensive index has a relatively good predetermined ratio. (For example, 80%) features can be selected as model-building features.

[Sixth Embodiment]
(Hardware of computer 1000)
FIG. 8 is a hardware diagram showing a configuration example of a computer that executes each process of the first to fifth embodiments of the present invention. For example, a feature quantity sorting device that executes the above-mentioned feature quantity sorting method and a health condition evaluation device that executes the above-mentioned health condition evaluation method are realized by the computer 1000.

The computer 1000 includes a processor 1001, a main storage device 1002, an auxiliary storage device 1003, a network interface 1004, an input device 1005, and an output device 1006, including a CPU connected to each other via an internal communication line 1009 such as a bus. It is a computer equipped.

The processor 1001 is a processor such as a microprocessor, a GPU (Graphics Processing Unit), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or any combination thereof. The processor 1001 controls the operation of the entire computer 1000. Further, the main storage device 1002 is composed of, for example, a volatile semiconductor memory and is used as a work memory of the processor 1001. The auxiliary storage device 1003 is composed of a large-capacity non-volatile storage device such as a hard disk device, an SSD (Solid State Drive), or a flash memory, and is used for holding various programs and data for a long period of time.

The executable program 1100 stored in the auxiliary storage device 1003 is loaded into the main storage device 1002 when the computer 1000 is started up or when necessary, and the processor 1001 executes the executable program 1100 loaded in the main storage device 1002. The above-mentioned device that executes various processes is realized.

The executable program 1100 may be recorded on a non-temporary recording medium, read from the non-temporary recording medium by a medium reading device, and loaded into the main storage device 1002. Alternatively, the executable program 1100 may be acquired from an external computer via a network and loaded into the main storage device 1002.

The network interface 1004 is an interface device for connecting the computer 1000 to each network in the system or communicating with other computers. The network interface 1004 is composed of, for example, a NIC (Network Interface Card) such as a wired LAN (Local Area Network) or a wireless LAN.

The input device 1005 is composed of a keyboard, a pointing device such as a mouse, and the like, and is used for a user to input various instructions and information to the computer 1000. The output device 1006 is composed of, for example, a display device such as a liquid crystal display or an organic EL (Electro Luminescence) display, and an audio output device such as a speaker, and is used to present necessary information to the user when necessary.

Although the present invention has been illustrated and described with reference to a preferred embodiment of the present invention, the present invention is not necessarily limited to the embodiment having all the configurations described in the present invention, and the present invention is not limited to the embodiment. Each embodiment can be combined with each other without departing from the spirit, or a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of one embodiment can be replaced with the configuration of another embodiment. It should be understood by those skilled in the art that other configurations can be added, deleted, or replaced with respect to a portion of the configurations of each embodiment.

For example, in the above, each embodiment has been described by taking the amount of wear of the tool as an example, but the present invention is not limited to this, and the feature amount selection method and the health condition evaluation method of the present invention are not only applied to the tool but also the bearing. It can also be used for equipment such as motors, blowers, machine tools, parts, tools (that is, objects to be measured) and the like. Taking a bearing as an example, detection information in the operating process of the bearing can be acquired by a vibration sensor, and for example, a feature amount including an amplitude value and a frequency spectrum of a vibration signal is acquired. After that, each feature amount is selected with the remaining life of the bearing as the health state, and finally, a bearing health state evaluation model for evaluating the health state with an arbitrary feature value is established.

1000: computer, 1001: processor, 1002: main memory

Claims (13)

In the feature amount selection method performed by the feature amount selection device that selects the feature amount suitable for evaluation of the health condition of the object to be measured from a plurality of feature amounts.
A data acquisition step in which a plurality of feature values of the object to be measured are acquired for each of the plurality of feature quantities, and the feature values are numerical values for expressing the feature quantity.
For each of the feature quantities waiting to be evaluated among the plurality of feature quantities, a non-monotonic index and a deviation index of the feature quantity are calculated based on the plurality of feature values acquired in the data acquisition step, and among them, the said The non-monotonicity index reflects the non-monotonicity of the feature value, and the deviation index reflects the degree of deviation of the feature value with respect to the overall change tendency, and the feature amount evaluation step.
For each of the feature quantities waiting to be evaluated, a comprehensive index is calculated based on the non-monotonic index and the deviation index, and it is suitable for evaluating the health condition of the object to be measured based on a preset index threshold value. The feature amount selection step to select the optimum feature amount, and
A feature quantity selection method characterized by containing. In the data acquisition step, a plurality of feature values of one object to be measured are acquired in chronological order for each of the plurality of feature quantities.
In the feature amount evaluation step, the non-monotonicity index and the deviation index are set as the first-class non-monotonicity index and the first non-monotonicity index, respectively, based on a plurality of feature values acquired in chronological order for one object to be measured in the data acquisition step. Calculated as a first-class deviation index,
The comprehensive index is calculated based on the first-class non-monotonic index and the first-class deviation index.
The index threshold includes at least one of a threshold set for the comprehensive index, a threshold set for the first-class non-monotonic index, and a threshold set for the first-class deviation index. The feature amount selection method according to claim 1. In the data acquisition step, for each of the plurality of feature quantities, a plurality of feature values of the subject to be measured and a corresponding health condition value are acquired.
In the feature amount evaluation step, the non-monotonicity index and the deviation index are set to the first type non-monotonic, respectively, based on the plurality of feature values acquired for one object to be measured and the corresponding health condition values in the data acquisition step. Calculated as a sex index and a first-class deviation index,
The comprehensive index is calculated based on the first-class non-monotonic index and the first-class deviation index.
The index threshold includes at least one of a threshold set for the comprehensive index, a threshold set for the first-class non-monotonic index, and a threshold set for the first-class deviation index. The feature amount selection method according to claim 1, wherein the feature amount is selected. In the data acquisition step, one or a plurality of feature values and corresponding health condition values are acquired for each of the plurality of objects to be measured of the same type for each of the plurality of feature quantities.
In the feature amount evaluation step, a non-monotonic index and a deviation index are based on one or more feature values and corresponding health condition values acquired for each of the plurality of objects to be measured of the same type in the data acquisition step. Are calculated as a second-class non-monotonic index and a second-class deviation index, respectively.
The comprehensive index is calculated based on the second-class non-monotonic index and the second-class deviation index.
The index threshold includes at least one of a threshold set for the comprehensive index, a threshold set for the second-class non-monotonic index, and a threshold set for the second-class deviation index. The feature amount selection method according to claim 1, wherein the feature amount is selected. In the data acquisition step, one or a plurality of feature values and corresponding health condition values are further acquired for each of the plurality of objects to be measured of the same type for each of the plurality of feature quantities.
In the feature amount evaluation step, a non-monotonic index and a non-monotonic index are further based on one or more feature values and corresponding health condition values acquired for each of the plurality of objects to be measured of the same type in the data acquisition step. The deviation index is calculated as a second-class non-monotonic index and a second-class deviation index, respectively.
The comprehensive index is calculated based on the first-class non-monotonic index, the first-class deviation index, the second-class non-monotonic index, and the second-class deviation index.
The index threshold is a threshold set for the comprehensive index, a threshold set for the first-class non-monotonic index, a threshold set for the first-class deviation index, and a second-class non-monotonicity. The feature amount selection method according to claim 2 or 3, wherein the threshold value set for the index and at least one of the threshold values set for the second type deviation index are included. In the feature amount evaluation step, the non-monotonic index and the deviation index of the feature amount are calculated according to the following steps for each of the feature amounts waiting to be evaluated.
Step 1. The feature values of the feature amount acquired in the data acquisition step are plotted in a coordinate system with the feature amount on the vertical axis and the number on the horizontal axis to obtain original data, and the numbers are in the time order for acquiring the feature values. Represents
Step 2, draw the upper and lower envelopes of the original data,
Step 3, calculate the center line based on the upper envelope and the lower envelope.
Step 4, based on the center line, calculate the non-monotonic index according to the formula (2) below, and calculate the deviation index according to the formula (3) below.

Here, n is the number of data points, y is the feature value of the i-th data point on the center line, and Di is the same lateral line as the i-th data point on the center line and the upper envelope or the lower envelope. The feature amount selection method according to claim 2, wherein the coordinate is a distance from a data point. In the feature amount evaluation step, the non-monotonic index and the deviation index of the feature amount are calculated according to the following steps for each of the feature amounts waiting to be evaluated.
Step 1. The feature value of the feature amount and the corresponding health state value acquired in the data acquisition step are plotted in a coordinate system with the feature amount on the vertical axis and the health state on the horizontal axis to obtain original data, and the feature value is obtained. The number of represents the time order to acquire the feature value,
Step 2, draw the upper and lower envelopes of the original data,
Step 3, calculate the center line based on the upper envelope and the lower envelope.
Step 4, based on the center line, calculate the non-monotonic index according to the formula (2) below, and calculate the deviation index according to the formula (3) below.

Here, n is the number of data points, y is the feature value of the i-th data point on the center line, and Di is the same lateral line as the i-th data point on the center line and the upper envelope or the lower envelope. The feature amount selection method according to claim 3 or 4, wherein the coordinate is a distance from a data point. In step 2, the upper envelope and the lower envelope of the original data are drawn according to the following five rules.
Rule 1: The first point and the last point of the original data belong to the upper envelope and the lower envelope at the same time.
Rule 2: For other data points except the first and last points of the original data, if the quadratic derivative with respect to the abscissa is zero, the data points are simultaneously the top envelope and the said. Belongs to the lower envelope,
Rule 3: For the second data point from the back, if the quadratic derivative for its abscissa is greater than zero, then the data point belongs to the lower envelope and if the quadratic derivative for its abscissa is less than zero. For example, the data point belongs to the upper envelope,
Rule 4: For the first point, the second from the back, and the other data points except the last point, if the quadratic derivative for the horizontal coordinate is not zero, then the quadratic derivative of the data point Calculate the product of the next data point with the quadratic derivative, and if the product is greater than zero, then the data point belongs to the upper and lower envelopes at the same time, and if the product is less than zero. And if the quadratic derivative of the data point is greater than zero, then the data point belongs to the lower envelope, and if the product is less than zero and the quadratic derivative of the data point is less than zero, then the data. The point belongs to the upper envelope and
Rule 5: When the data points of the original data corresponding to some horizontal coordinates are assigned only to the upper envelope or the lower envelope and the data points of the other envelope are lost, before and after the data points. The feature amount selection method according to claim 7, wherein the data points of the envelope are complemented by performing linear interpolation using the numerical values of the points adjacent to the envelope. After the data acquisition step, the feature amount simplification step before the feature amount evaluation step is further included, and in the feature amount simplification step, the plurality of feature amounts are arbitrarily set as the feature amount awaiting evaluation. For the two features, the linear correlation coefficient between the two features is calculated based on the plurality of feature values acquired in the data acquisition step and the corresponding health condition values, and the calculated linear correlation coefficient is When the preset correlation threshold is exceeded, it is determined that the two feature quantities are substantially the same, and the calculation time of the feature value among the two feature quantities is relatively long from the feature quantity awaiting evaluation. The feature amount selection method according to claim 3 or 4, wherein a long feature amount is deleted. In the data acquisition step, sensor signal values are acquired from each sensor for detecting the feature amount installed on the object to be measured, and the acquired sensor signal values are preprocessed and then preprocessed. The feature amount selection method according to claim 1, wherein the feature value is extracted from a later sensor signal value. The feature amount selection method according to claim 1, wherein the health condition is any one of cumulative operating time, remaining life, and wear amount. In the health condition evaluation method executed by the health condition evaluation device that evaluates the health condition of the object to be measured,
Each step of the feature quantity selection method according to any one of claims 3 to 11 and
A part or all of the feature amount is selected as the model building feature amount from the optimum feature amount, and based on the feature value acquired for each model building feature amount in the data acquisition step and the corresponding health condition value. , A model building step to build a mathematical model between the model building features and health status,
A health condition evaluation step for detecting the object to be measured and evaluating the health condition of the object to be measured based on the feature value of the detected feature amount for model construction and the mathematical model constructed in the model construction step.
A health condition assessment method characterized by inclusion. The model construction step is characterized in that by sorting the optimum feature amount according to the superiority or inferiority of the comprehensive index, a predetermined ratio of feature amounts having a relatively good comprehensive index is selected as the model construction feature amount. The health condition evaluation method according to claim 12.

Images