THRESHOLD CALCULATION DEVICE AND STORAGE MEDIUM

Abstract

A threshold value calculation device which calculates a threshold value for detecting abnormalities occurring in a factory. The threshold value calculation device: uses input signals input from sensors in the factory to calculate normal data from a normal model created on the basis of input signals in a normal state; calculates a feature quantity of an input signal to be diagnosed for abnormality; compares the deviation between the normal data and the feature quantity with a threshold value to diagnose the input signal for abnormality; and, if the diagnostic result is positive, receives the result of a confirmatory test for determining whether the diagnostic result is a false positive or true positive, and adjusts the threshold value on the basis of the frequency of occurrence of false positives.

Description

[Technical field]

[0001] The present invention relates to a threshold calculation device and a storage medium. [Background technology]

[0002] Conventionally, there is a technology for diagnosing abnormalities in machines. In the abnormality diagnosis, a threshold value may be used. For example, a value related to the diagnosis target, or a value calculated from the value or the like using a predetermined calculation formula, is compared with a threshold value, and it is possible to diagnose whether the value is abnormal or normal depending on whether the value is equal to or greater than the threshold value used for diagnosing normality and abnormality, or whether the value is equal to or less than the threshold value (for example, see Non-Patent Document 1).

[0003] This threshold is set by comparing a set of normal values ​​with a set of abnormal values, and setting a value that can classify most of these sets as the threshold. For example, there is a method of setting the threshold so that the product of the false alarm rate p, which is the probability of a normal value being erroneously diagnosed as an abnormal value, and the loss Cp when a false alarm occurs in which a normal value is erroneously diagnosed as an abnormal value, is equal to the product of the undetected rate q, which is the probability of an abnormality not being detected because an abnormal value is erroneously determined to be normal, and the loss Cq when an undetected occurs (see, for example, Non-Patent Document 2).

[0004] In addition, there is a clustering analysis as a method for classifying input data into multiple clusters based on data patterns. One of the abnormality diagnosis methods in clustering analysis is to use the distance from the center of gravity of the nearest cluster. A threshold value is used for the abnormality diagnosis (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2017-33471 A [Non-patent literature]

[0006] [Non-Patent Document 1] Genichi Taguchi (ed.), Society of Quality Engineering (ed.), "Quality Engineering Handbook," Nikkan Kogyo Shimbun, 2007, pp. 379-384 [Non-Patent Document 2] Hiroshi Yano (ed.), "Genichi Taguchi Essay Collection, Vol. 4," Japanese Standards Association, 2012, pp. 393-399 Summary of the Invention [Problem to be solved by the invention]

[0007] In the above-mentioned example of anomaly detection, a model of normal data is created based on normal data, and if the deviation from the normal model exceeds a predetermined threshold, it is diagnosed as an anomaly. In anomaly detection, false positives occur, where an abnormality is diagnosed when it is not, and false negatives occur, where an abnormality is diagnosed when it is not. False negatives can lead to breakdowns or the release of defective products.

[0008] If the threshold is set too tightly to prevent false negatives, the rate of false positives increases. In actual operation, workers check whether a positive diagnosis is actually positive. As the number of false positives increases, the cost of checking increases.

[0009] In the field of manufacturing anomaly diagnosis, technology for calculating appropriate thresholds is desired. [Means for solving the problem]

[0010] A threshold calculation device according to one aspect of the present disclosure is a threshold calculation device that calculates a threshold for detecting anomalies that occur in a factory, and includes an anomaly diagnosis unit that uses an input signal input from a sensor present in the factory to calculate normal data from a normal model created based on the input signal under normal conditions, calculates a feature amount of the input signal to be diagnosed as an anomaly, and diagnoses an anomaly in the input signal by comparing the deviation between the normal data and the feature amount with a threshold, a confirmation result acquisition unit that, if the diagnosis result by the anomaly diagnosis unit is positive, receives the result of a confirmation test to determine whether the diagnosis result is a false positive or a true positive, and a threshold adjustment unit that adjusts the threshold based on the frequency of occurrence of false positives. a normal model storage unit that stores a plurality of the normal models; and a model switching unit that monitors changes in a manufacturing state of a product in a factory, selects a normal model suitable for the manufacturing state of the factory from among the plurality of normal models, and switches the normal model; Equipped with. A computer-readable storage medium according to one aspect of the present disclosure uses an input signal input from a sensor present in a factory, calculates normal data from a normal model created based on the input signal under normal conditions, calculates a feature amount of the input signal to be diagnosed as an anomaly, compares the deviation between the normal data and the feature amount with a threshold value, and diagnoses an anomaly in the input signal. If the diagnosis result is positive, the computer-readable storage medium receives the result of a confirmation test as to whether the diagnosis result is a false positive or a true positive, and adjusts the threshold value based on the frequency of occurrence of false positives. The change in the manufacturing state of the product in the factory is monitored, and a normal model suitable for the manufacturing state of the factory is selected from among a plurality of normal models prepared in advance, and the normal model is switched. . Effect of the Invention

[0011] According to one aspect of the present invention, it is possible to calculate an appropriate threshold value for diagnosing an abnormality in a machine. [Brief description of the drawings]

[0012] [Figure 1] FIG. 2 is a hardware configuration diagram of a threshold calculation device. [Diagram 2] 1 is a block diagram of a threshold calculation device according to a first disclosure. [Diagram 3] 1 is a table showing the relationship between true positives, false positives, false negatives, and true negatives. [Figure 4] FIG. 1 is a diagram showing the relationship between true negatives and false positives. [Diagram 5] FIG. 1 is a diagram showing the relationship between true positives and false negatives. [Figure 6] FIG. 13 shows the frequency of positive occurrence before threshold adjustment. [Figure 7] FIG. 13 shows the frequency of positive occurrence before threshold adjustment. [Figure 8] 13 is a flowchart illustrating a process of a first threshold calculation device. [Figure 9] FIG. 13 is a block diagram of a threshold calculation device according to a second disclosure. [Figure 10] FIG. 13 is a diagram showing changes in threshold values ​​with respect to manufacturing conditions. [Figure 11] 13 is a flowchart showing a process of a threshold calculation device according to a second disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] [First Disclosure] The threshold calculation device 100 calculates thresholds for detecting abnormalities in a factory, machines and equipment in the factory, and products manufactured in the factory. The hardware configuration of a threshold calculation device 100 according to the present disclosure will be described with reference to FIG. The threshold calculation device 100 is an information processing device such as an abnormality detection device for factory facilities and machines, a PC (personal computer) that monitors the state of a factory, or a numerical control device that monitors the state of a machine tool. The CPU 111 included in the threshold calculation device 100 is a processor that controls the entire threshold calculation device 100. The CPU 111 reads a system program processed in the ROM 112 via a bus, and controls the entire threshold calculation device 100 according to the system program. The RAM 113 temporarily stores temporary calculation data, display data, various data input by the user via the input unit 71, and the like.

[0014] The display unit 70 is a monitor or the like attached to the threshold calculation device 100. The display unit 70 displays an operation screen, a setting screen, and the like of the threshold calculation device 100.

[0015] The input unit 71 is a keyboard, a touch panel, or the like that is integrated with the display unit 70 or is separate from the display unit 70. A user operates the input unit 71 to input data to a screen displayed on the display unit 70. The display unit 70 and the input unit 71 may be a mobile terminal.

[0016] The non-volatile memory 114 is a memory that maintains its stored state even when the power supply of the threshold calculation device 100 is turned off, for example, by being backed up by a battery (not shown). The non-volatile memory 114 stores programs read from an external device via an interface (not shown), programs input via the input unit 71, and various data acquired from each unit of the threshold calculation device 100 and sensors in a factory. The programs and various data stored in the non-volatile memory 114 may be expanded in the RAM 113 when executed / used. In addition, various system programs are written in the ROM 112 in advance.

[0017] FIG. 2 is a block diagram of the threshold calculation device 100 of the first disclosure. The threshold calculation device 100 includes a normal model storage unit 11, an abnormality diagnosis unit 12, a confirmation result acquisition unit 13, and a threshold adjustment unit . The normal model storage unit 11 creates a normal model based on data in normal times (when the mechanical equipment starts operating, in a stable state, and in a steady state). In detecting an abnormality, an abnormality in the input signal is diagnosed based on the deviation between normal data obtained from the normal model and a feature amount obtained from the input signal. Methods for analyzing an abnormality include, but are not limited to, the MT method.

[0018] The abnormality diagnosis unit 12 performs automatic abnormality inspection. The abnormality diagnosis unit 12 acquires input signals such as sensor signals, and calculates the feature quantities of the input signals. The input signals include the temperature, vibration, operating time, torque, power consumption, and ambient temperature of the machine. The abnormality diagnosis unit 12 calculates a scalar value indicating the degree of abnormality from the deviation between the feature quantities of the input signals and normal data obtained from the normal model. The abnormality diagnosis unit 12 compares the scalar value with a threshold value, and diagnoses an abnormality when the scalar value exceeds the threshold value.

[0019] The confirmation result acquisition unit 13 receives the results of a confirmation test for determining whether the "positive diagnosis" was successful for an input signal diagnosed as "positive" by an automatic test. One type of confirmation test is a manual test performed by a product management personnel. The method of the confirmation test is not limited to this. The diagnosis results are based on the results of the automatic test and the confirmation test, and can be one of the following, as shown in Figure 3: "True positive: automatic test (positive) and true state (positive)", "False positive: automatic test (positive) and true state (negative)", "False negative: automatic test (negative) and true state (positive)", or "True negative: automatic test (negative) and true state (negative)". In this disclosure, the true state of a product diagnosed as "negative" by automatic testing is all considered to be "true negative" and the threshold value described below is calculated.

[0020] The relationship between "true positive", "false positive", "true negative" and "false negative" will be explained with reference to the graphs of FIG. 4 and FIG. The horizontal axis of the graphs in Figures 4 and 5 is a feature value indicating the degree of abnormality of a product. The vertical axis of the graph in Figure 4 is the probability of a "negative" result, and the vertical axis of the graph in Figure 5 is the probability of a "positive" result.

[0021] FIG. 4 shows the probability density distribution of "negative". If the feature is smaller than the threshold, the result of the automated test is "negative". In this disclosure, if the automated test is "negative", it is considered a "true negative". If the feature is larger than the threshold, the result of the automated test is "positive". If the result of the automated test is "positive" and the result of the confirmatory test is "negative", it is a "false positive".

[0022] Figure 5 shows the probability density distribution of "positives". Since the number of "positives" is small, the graph of the probability density distribution of "positives" is actually an unknown graph. However, since it can be assumed that the higher the degree of abnormality, the higher the probability of a "positive" occurring, the probability density distribution of "positives" has a peak to the right of the probability density distribution of "negatives". As shown in Figure 5, the probability density distribution of “positives” is also classified into “true positives” and “false negatives” depending on the threshold value.

[0023] The smaller the threshold, the higher the probability that the automated test will diagnose "positive: abnormality present." The larger the threshold, the higher the probability that the automated test will diagnose "negative: no abnormality." The smaller the threshold, the stricter the diagnosis. That is, setting a strict threshold increases the probability of a "false positive" and decreases the probability of a "false negative." Setting a loose threshold decreases the probability of a "false positive" and increases the probability of a "false negative." There is a relationship between the probability of a "false positive" and the probability of a "false negative." In the present disclosure, the frequency of occurrence of "false positives" and "false negatives" is adjusted using the relationship between the probabilities of "false positives" and "false negatives."

[0024] In a "false negative," a product is diagnosed as normal even though it actually has an abnormality. Ignoring an abnormality or releasing a product with an abnormality can lead to machine breakdowns and the outflow of defective products, so "false negatives" must be prevented. However, setting a strict threshold to prevent "false negatives" increases the probability of "false positives." As the number of "false positives" increases, confirmatory testing is required even when the product is actually normal, which takes time and effort. By reducing the number of "false positives" while tolerating a certain degree of "false negatives," it is possible to prevent machine breakdowns and the release of defective products and reduce the effort required for confirmation testing.

[0025] In adjusting the threshold, the threshold is adjusted so that the ratio of "false positives" to "negatives" is as small as possible within a range greater than x%. Here, (I) the number of "negatives" can be either the number of products diagnosed as "negative" in the automated test, or (II) the number of products diagnosed as "negative" in the automated test plus the number of "false negative" products. The value of x is set by the user.

[0026] In the present disclosure, the threshold is adjusted using the ratio of "false positives" to "negatives" to adjust the probability of "false negatives" and "false positives" occurring, thereby reducing the number of "false positives" while keeping the probability of "false negatives" within an acceptable range.

[0027] The threshold adjustment unit adjusts the threshold based on the results of automatic testing and confirmation testing. If "false positives" occur frequently, the threshold value is loosened. If "false positives" do not occur, the threshold value is tightened.

[0028] A method for setting the threshold value of the present disclosure will be described with reference to FIG. 6. The threshold value can be manually set by an operator. Preferably, the initial value of the threshold value is set so that the frequency (probability) of "false positive" is high. When the threshold value is set in this way, "positive" alerts are frequently generated as shown in FIG. 6. If the cause of the alert is a "false positive", the threshold value is adjusted to lower the probability of "false positive", but the probability of "false positive" is adjusted so that it is not smaller than x%. If the initial value of the threshold value is set so that the probability of "false positive" is initially high, the probability of "false negative" is low. By adjusting the threshold value, the probability of "false negative" gradually increases, but the probability of "false negative" is ultimately suppressed within an acceptable range, so that the threshold value can be adjusted while suppressing the probability of "false negative" within an acceptable range. For example, when threshold adjustment begins, 28 out of 30 detected "positives" are "false positives," but as adjustment continues, the probability of "false positives" decreases to 24, 20, .... Eventually, for example, out of 30 detected "positives," 5 will be "false positives" and the probability will stabilize (see Figure 7).

[0029] The process of the first threshold calculation device 100 will be described with reference to the flowchart of FIG. The threshold calculation device 100 first sets a threshold (step S1). The threshold calculation device 100 performs an automatic inspection of an input signal (step S2). If the automatic inspection results in a diagnosis of "negative" (step S3; negative), the threshold calculation device 100 diagnoses the product as "negative" (step S4).

[0030] If the input signal is judged as "positive" as a result of the automatic test (step S3; positive), a confirmation test of the input signal is performed (step S5). If the result of the confirmation test is diagnosed as negative (step S6; negative), the threshold calculation device 100 diagnoses the confirmation result as "false positive" (step S7). If the result of the confirmation test is a positive diagnosis (step S6; positive), the threshold calculation device 100 diagnoses the confirmation result as a "true positive" (step S8).

[0031] The threshold calculation device 100 calculates the ratio of "false positives" to the number of "negatives" based on the results of the automatic test and the confirmatory test (step S9), and adjusts the threshold so that this ratio is smaller than x% (step S10). If the threshold is stabilized as a result of the adjustment (step S11; YES), the threshold calculation device 100 ends the calculation of the threshold. If the threshold is not stabilized (step S11; NO), the threshold calculation device 100 proceeds to step S2 and continues adjusting the threshold.

[0032] As described above, the threshold calculation device 100 of the first disclosure calculates normal data from a normal model, calculates a scalar value that is a feature of the degree of abnormality from an input signal, and compares the scalar value with a threshold to diagnose the abnormality as "positive" or "negative." The initial value of the threshold is set to a strict value, and the threshold is adjusted based on the proportion of "false positives" among "negatives." The threshold calculation device 100 adjusts the threshold to reduce the probability of occurrence of a "false negative" as much as possible.

[0033] [Second Disclosure] The second disclosure will be described. The threshold calculation device of the second disclosure calculates a threshold for detecting an abnormality in a machine tool. The threshold calculation device 100 of the second disclosure may calculate a threshold for detecting an abnormality in a factory, a machine or equipment in the factory, or a product manufactured in the factory, similar to the threshold calculation device 100 of the first disclosure.

[0034] FIG. 9 is a block diagram of a threshold calculation device 100 according to the second disclosure. The threshold calculation device 100 according to the second embodiment includes a model switching unit 15. The model switching unit 15 selects an appropriate normal model from a plurality of normal models. The threshold calculation device 100 of the second disclosure stores the chatter vibration occurrence limit as a threshold value and detects chatter vibration. In order to detect chatter vibration abnormality, a vibration or torque signal is input as an input signal. The threshold calculation device 100 can also be applied to abnormality detection other than chatter vibration. Note that a description of parts that overlap with the threshold calculation device 100 of the first disclosure will be omitted.

[0035] The model switching unit 15 monitors changes in the manufacturing state of the product. The changes in the manufacturing state include changes in the input signal, changes in the analysis method and the feature extraction method, and changes in the workpiece, tools, processing method, and operating environment. FIG. 10 shows the variation of the threshold with respect to manufacturing conditions. Figure 10 shows the range of thresholds calculated in six different states. The six states are (1) the threshold when analyzing the presence or absence of abnormalities when performing end milling with machine 1 using the FFT (Fast Fourier Transform) method, (2) the threshold when analyzing the presence or absence of abnormalities when performing end milling with machine 1 using the P2P (Peak to Peak) method, (3) the threshold when analyzing the presence or absence of abnormalities when performing end milling with machine 2 using the FFT method, (4) the threshold when analyzing the presence or absence of abnormalities when performing end milling with machine 2 using the P2P method, (5) the threshold when analyzing the presence or absence of abnormalities when performing milling with machine 3 using the FFT method, and (6) the threshold when analyzing the presence or absence of abnormalities when performing milling with machine 3 using the P2P method.

[0036] In the case of (1), misdiagnosis will not occur if the threshold is set between about 5 and 9. The threshold range for (2) is about 7 to 11, the threshold range for (3) is about 2 to 4, the threshold range for (4) is about 8 to 10, the threshold range for (5) is about 2 to 7, and the threshold range for (6) is about 7.5 to 8. The area between the auxiliary lines in Figure 10 is a common threshold area. An appropriate threshold can be selected from the common area.

[0037] The process of the threshold calculation device 100 of the second disclosure will be described with reference to the flowchart in Fig. 11. The threshold calculation device 100 selects a normal model based on the current state (step S21). The threshold calculation device 100 performs automatic adjustment of the threshold (step S22). Step S22 corresponds to steps S1 to S9 in the first embodiment.

[0038] The threshold calculation device 100 monitors whether or not there is a state change. If a state change occurs (step S23; Yes), the threshold calculation device 100 switches the normal model corresponding to the current state to the normal model (step S24) and proceeds to step S22. If there is already an adjusted threshold, the threshold is also switched.

[0039] If no state change has occurred (step S23; No), the threshold calculation device 100 diagnoses whether the threshold has stabilized. If the threshold has stabilized (step S25; YES), the threshold calculation device 100 ends the calculation of the threshold. If the threshold has not stabilized (step S25; No), the threshold calculation device 100 proceeds to step S22 and continues adjusting the threshold.

[0040] As described above, the threshold calculation device 100 of the second disclosure stores multiple normal models and switches between normal models depending on state changes. It is possible to adjust the threshold depending on the type of machine, input signal, analysis method, feature amount, workpiece, tool, processing method, operating environment, etc. [Explanation of symbols]

[0041] 100 Threshold calculation device 11 Normal model memory section 12 Abnormality diagnosis section 13. Confirmation result acquisition section 14 Threshold adjustment section 15 Model switching section 111 CPU 112 ROM 113 RAM 114 Non-volatile memory

Claims

1. A threshold calculation device for calculating a threshold for detecting an abnormality occurring in a factory, an abnormality diagnosis unit that uses an input signal input from a sensor present in the factory, calculates normal data from a normal model created based on the input signal under normal conditions, calculates a feature amount of the input signal to be diagnosed as an abnormality, and diagnoses an abnormality in the input signal by comparing a deviation between the normal data and the feature amount and a threshold value; a confirmation result acquisition unit that, when a diagnosis result by the abnormality diagnosis unit is positive, receives a result of a confirmation test to determine whether the diagnosis result is a false positive or a true positive; a threshold adjustment unit that adjusts the threshold based on the frequency of occurrence of the false positives; a normal model storage unit that stores a plurality of the normal models; a model switching unit that monitors changes in a manufacturing state of a product in a factory, selects a normal model suitable for the manufacturing state of the factory from among the plurality of normal models, and switches the normal model; A threshold calculation device comprising:

2. The threshold calculation device according to claim 1 , wherein the frequency of occurrence of false positives is a ratio of the false positives to the negatives diagnosed by the abnormality diagnosis unit.

3. 2 . The threshold calculation device according to claim 1 , wherein the frequency of occurrence of false positives is a ratio of the false positives to a sum of the negatives diagnosed by the abnormality diagnosis unit and the false positives acquired by the confirmation result acquisition unit.

4. The threshold calculation device according to claim 2 , wherein the threshold adjustment unit adjusts the threshold so that the ratio is smaller than a predetermined value.

5. The normal model storage unit stores an adjusted threshold value, 2. The threshold calculation device according to claim 1, wherein the model switching unit switches the adjusted threshold to a normal model suited to a state of the factory.

6. When executed by one or more processors, Using an input signal input from a sensor present in a factory, normal data is calculated from a normal model created based on the input signal under normal conditions, a feature amount of the input signal to be subject to an abnormality diagnosis is calculated, and an abnormality in the input signal is diagnosed by comparing the deviation between the normal data and the feature amount with a threshold value; If the diagnosis is positive, accept the result of a confirmatory test to determine whether the diagnosis is false positive or true positive; adjusting the threshold based on the frequency of occurrence of the false positives; monitoring a change in a manufacturing state of a product in a factory, selecting a normal model suitable for the manufacturing state of the factory from among a plurality of normal models prepared in advance, and switching the normal model; A storage medium that stores computer-readable instructions.

Images