Trigger condition determination method for time series signal, abnormality diagnosis method for monitored equipment, and trigger condition determination device for time series signal
By generating decision tree models through machine learning and automatically determining triggering conditions, the inefficiency caused by manual intervention in existing technologies is solved, and automated anomaly diagnosis of complex equipment is achieved.
Patent Information
- Application Number
- CN202080107703.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-08
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2040-12-08
AI Technical Summary
Existing anomaly diagnosis methods require manual intervention to determine triggering conditions, resulting in low efficiency and time consumption, making them difficult to automate in complex equipment.
The system generates a decision tree model through machine learning, which automatically determines the triggering conditions. It utilizes the correlation coefficient and label data of the monitored object signal and the trigger candidate signal to achieve automated triggering condition determination.
It enables automatic determination of triggering conditions in complex equipment, simplifies the preparation process for anomaly diagnosis, and improves efficiency and accuracy.
Smart Images

Figure CN116601576B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a trigger condition determination method for a time series signal, an abnormality diagnosis method for a monitored device, and a trigger condition determination device for a time series signal. Background Art
[0002] In production equipment in factories and other places and experimental equipment in research institutes and other places, as one of the methods for performing abnormal diagnosis of equipment, there is a method such as the following. First, a large amount of time series signal data (hereinafter referred to as "monitoring object signals") representing the state of the monitored equipment (hereinafter referred to as "monitoring object equipment") when it is operating normally is collected, and these monitoring object signals are used to create a model of the monitored equipment when it is operating normally. In addition, when performing abnormal diagnosis of the monitored equipment, the distance from the above-mentioned model is calculated, and when the distance exceeds a predetermined threshold, it is determined to be abnormal. For example, Patent Document 1 discloses an abnormality diagnosis method that uses a principal component plane for the above-mentioned model.
[0003] Patent Document 1: International Publication No. 2017 / 104305
[0004] In the conventional abnormality diagnosis method represented by Patent Document 1, abnormality diagnosis is specifically performed through the following steps.
[0005] (1) Extract a time series signal of N points from the collected monitoring object signal.
[0006] (2) The extracted time series signal is represented as a point in N-dimensional space to create a benchmark.
[0007] (3) For the signal to be diagnosed as abnormality, the Q statistic (the distance from the principal component plane) is calculated.
[0008] (4) When the Q statistic exceeds a predetermined threshold, it is determined to be abnormal.
[0009] In the conventional abnormality diagnosis method, as shown in (1) above, the monitoring target interval is cut out from the monitoring target signal, and a model of the normal operation of the monitoring target device is created. In addition, when performing abnormality diagnosis of the monitoring target device, the monitoring target interval is also cut out from the monitoring target signal, and the distance to the model created from the monitoring target signal during normal operation is calculated. For simplicity, the condition that specifies the timing of cutting out the monitoring target interval from the monitoring target signal is called a "trigger condition", and the time series signal that can become the trigger condition is called a "trigger candidate signal".
[0010] In order to create an appropriate model when the monitored device is operating normally, when monitoring target intervals are cut out from multiple monitoring target signals collected from the monitored device and overlapped, the cut waveforms must overlap to a certain extent. In the existing abnormality diagnosis method, the overlap of the waveforms of the monitoring target interval is specifically performed through the following steps.
[0011] (1) Using a trigger condition, a monitoring target section is extracted from each of a plurality of monitoring target signals.
[0012] (2) The horizontal axis of each monitoring target section is transformed (for example, the time on the horizontal axis is transformed into a crank angle, etc.).
[0013] (3) The vertical axis of each monitoring target section is transformed (for example, normalized).
[0014] In the conventional abnormality diagnosis method, the trigger condition (1) is determined manually by comparing the monitored signal with the trigger candidate signal to overlap the cut waveforms, for example. Therefore, it takes effort and time to determine the trigger condition. Summary of the invention
[0015] The present invention is completed in view of the above content, and its purpose is to provide a trigger condition determination method for a time series signal that can automatically determine the trigger condition for intercepting a monitoring object interval from a monitoring object signal, an abnormality diagnosis method for a monitoring object device, and a trigger condition determination device for a time series signal.
[0016] In order to solve the above-mentioned problems and achieve the purpose, when performing abnormality diagnosis of a monitored device, the method for determining a trigger condition of a time series signal involved in the present invention determines a condition, i.e., a trigger condition, for extracting a monitored object interval that becomes the object of the above-mentioned abnormality diagnosis from a time series signal, i.e., a monitored object signal, indicating the state of the above-mentioned monitored device, and is characterized in that it includes: a collecting step, in which a signal group consisting of one or more monitored object signals related to the above-mentioned monitored device and a trigger candidate signal is collected, and the above-mentioned trigger candidate signal represents a time series signal that can become the above-mentioned trigger condition among the time series signals related to the above-mentioned monitored device and detected at the same time as the above-mentioned monitored object signal; An interception process, in which, for the above-mentioned signal group, the monitoring object interval of the above-mentioned monitoring object signal is intercepted based on a specified benchmark; a model generation process, in which, for the above-mentioned signal group, the start time of the intercepted above-mentioned monitoring object interval is determined, and label data for turning on (ON) the label at the start time and turning off (OFF) at other times is generated, and a learning model is generated through machine learning that takes one or more of the above-mentioned trigger candidate signals at each time as input and takes the above-mentioned label data at each time as output; and a trigger condition determination process, in which, for the monitoring object signal for performing the above-mentioned abnormal diagnosis, the above-mentioned trigger condition is determined using the above-mentioned learning model.
[0017] In addition, the trigger condition determination method of the time series signal involved in the present invention is completed on the basis of the above-mentioned invention, and is characterized in that, for the first monitoring object signal selected from the multiple monitoring object signals collected in the above-mentioned collection process, the above-mentioned interception process intercepts the monitoring object interval based on the device characteristics of the above-mentioned monitoring object device, and for the monitoring object signals other than the above-mentioned first monitoring object signal among the above-mentioned multiple monitoring object signals, the above-mentioned interception process searches for the interval with the largest correlation coefficient with the waveform contained in the monitoring object interval of the above-mentioned first monitoring object signal, thereby intercepting the monitoring object interval respectively.
[0018] In addition, the method for determining trigger conditions of a time series signal involved in the present invention is completed on the basis of the above invention, and is characterized in that the above learning model is a decision tree.
[0019] In addition, the trigger condition determination method of the time series signal involved in the present invention is completed on the basis of the above-mentioned invention, and is characterized in that, in the above-mentioned model generation process, when the above-mentioned trigger candidate signal is a single pulse signal, machine learning is performed after the above-mentioned trigger candidate signal is converted into a sawtooth wave.
[0020] In addition, the trigger condition determination method of the time series signal involved in the present invention is completed on the basis of the above-mentioned invention, and is characterized in that when the discrimination accuracy cannot obtain the specified value during machine learning in the above-mentioned model generation process, it returns to the above-mentioned interception process, shifts the monitoring object interval intercepted last time forward and backward, and after re-intercepting the monitoring object interval of the above-mentioned monitoring object signal, the above-mentioned model generation process is performed again.
[0021] In order to solve the above-mentioned problems and achieve the purpose, the abnormality diagnosis method of the monitored object device involved in the present invention is characterized in that, according to the trigger condition determined by the trigger condition determination method of the above-mentioned time series signal, the signal of the monitored object interval that becomes the object of the above-mentioned abnormality diagnosis is intercepted and accumulated from the time series signal representing the state of the monitored object device, that is, the monitored object signal, and the abnormality diagnosis of the above-mentioned monitored object device is performed based on the accumulated signal.
[0022] In order to solve the above-mentioned problems and achieve the purpose, the trigger condition determination device of the time series signal involved in the present invention determines the condition, i.e., the trigger condition, for intercepting the monitoring object interval that becomes the object of the above-mentioned abnormal diagnosis from the time series signal, i.e., the monitoring object signal, representing the state of the above-mentioned monitoring object device, when performing abnormal diagnosis of the above-mentioned monitoring object device, and is characterized in that it includes: a collection unit, which collects a signal group consisting of one or more monitoring object signals related to the above-mentioned monitoring object device and a trigger candidate signal, the above-mentioned trigger candidate signal representing the time series signal related to the above-mentioned monitoring object device and detected at the same time as the above-mentioned monitoring object signal, which can become the above-mentioned trigger a time series signal of a triggering condition; an interception unit, for the above-mentioned signal group, intercepts the monitoring object interval of the above-mentioned monitoring object signal based on a specified benchmark; a model generation unit, for the above-mentioned signal group, determines the starting time of the intercepted monitoring object interval, generates label data for turning on (ON) the label at the starting time and turning off (OFF) the label at other times, and generates a learning model through machine learning that takes one or more of the above-mentioned trigger candidate signals at each time as input and takes the above-mentioned label data at each time as output; and a trigger condition determination unit, for the monitoring object signal for performing the above-mentioned abnormal diagnosis, uses the above-mentioned learning model to determine the above-mentioned trigger condition.
[0023] According to the present invention, by using a learning model that has learned the conditions under which a monitoring target signal and a trigger candidate signal are triggered to be turned on, it is possible to automatically determine a trigger condition for cutting out a monitoring target section from the monitoring target signal. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 This is a block diagram showing a schematic configuration of a trigger condition determination device for a time series signal according to an embodiment of the present invention.
[0025] Figure 2 This is a flowchart showing the flow of a method for determining a trigger condition of a time-series signal according to an embodiment of the present invention.
[0026] Figure 3 It is a diagram schematically showing the contents of the interception process of the method for determining the trigger condition of the time series signal according to the embodiment of the present invention.
[0027] Figure 4 It is a diagram schematically showing the contents of the interception process of the method for determining the trigger condition of the time series signal according to the embodiment of the present invention.
[0028] Figure 5 It is a diagram schematically showing the content of the model generation step of the method for determining the trigger condition of the time series signal according to the embodiment of the present invention.
[0029] Figure 6 This is a diagram schematically showing a decision tree generated in the model generation step of the method for determining trigger conditions for a time series signal according to the embodiment of the present invention.
[0030] Figure 7 This is a diagram schematically showing how a single pulse signal is converted into a sawtooth wave in a model generation step of a method for determining a trigger condition for a time series signal according to an embodiment of the present invention.
[0031] Figure 8 This is a diagram schematically showing a sawtooth wave converted in the model generation step of the method for determining a trigger condition of a time series signal according to the embodiment of the present invention.
[0032] Fig. 9 This is a diagram showing a monitoring target signal and a trigger candidate signal in an example of a method for determining a trigger condition for a time series signal according to an embodiment of the present invention.
[0033] Fig.10 This is a diagram showing a state where the waveforms of the cut-out monitoring target sections are overlapped in an example of the method for determining the trigger condition of the time series signal according to the embodiment of the present invention.
[0034] Fig.11 This is a diagram showing a method of determining a trigger condition using a decision tree in an example of a method of determining a trigger condition for a time series signal according to an embodiment of the present invention.
[0035] Fig.12 This is a diagram showing a state where the waveforms of the cut-out monitoring target sections are overlapped in an example of the method for determining the trigger condition of the time series signal according to the embodiment of the present invention. DETAILED DESCRIPTION
[0036] A method for determining a trigger condition of a time series signal, a method for diagnosing an abnormality of a monitored device, and a device for determining a trigger condition of a time series signal (hereinafter referred to as a “learning device”) according to an embodiment of the present invention will be described with reference to the accompanying drawings.
[0037] (Trigger condition determination device)
[0038] The trigger condition determination device is a device that determines a trigger condition, which is a condition for cutting out a monitoring target section from a monitoring target signal, when abnormality diagnosis of monitoring target equipment is performed in production equipment such as a factory or experimental equipment such as a research institute.
[0039] Here, as described above, the monitoring object signal shows a time series signal indicating the state of the monitoring object signal. The monitoring object signal varies depending on the type of the monitoring object device. For example, when the monitoring object device is a "motor", the motor's current, speed, etc. are used as the monitoring object signal. In the trigger condition determination device, one or more monitoring object signals and corresponding trigger candidate signals are always collected. The trigger candidate signal represents a time series signal that is related to the monitoring object device and is detected at the same time as the monitoring object signal.
[0040] Usually, when one or more conditions are met, the device to be monitored starts to operate. The trigger candidate signal refers to the condition, such as a signal of turning on (ON), turning off (OFF), etc. If the trigger condition is known in advance, it is easy to intercept the monitoring object interval. However, when there are a large number of various devices and they operate in a complex manner, it is sometimes difficult to easily determine the trigger condition. Alternatively, there are also cases where the trigger candidate signal that directly represents the trigger condition is not necessarily taken into the database and only an indirect form of the signal exists.
[0041] Therefore, a signal group that specifies the equipment action conditions is selected as a candidate trigger signal, and the rules for starting the action are extracted from the history of these signals through machine learning, etc., to determine the conditions for intercepting the monitoring object interval. In addition, the candidate trigger signal is preferably a signal that has no direct relationship with the abnormality of the object process, equipment, etc., and the signal indicating the opening (ON) / closing (OFF) of various instruction values and specific events is a candidate. However, the candidate trigger signal can also include the monitoring object signal itself.
[0042] In addition, the monitoring object interval represents an interval intercepted for abnormal diagnosis of the monitoring object signal. When intercepting the monitoring object interval from the monitoring object signal, the time to start intercepting the monitoring object signal (hereinafter referred to as the "interception start time") and the interception width are specified. At this time, the interception start time is specified based on the value of the trigger candidate signal collected simultaneously with the monitoring object signal. In addition, the interception width varies depending on the type of the monitoring object device. For example, when the monitoring object device is a "motor", the interval of motor acceleration can also be specified as the interception width. Alternatively, if it is a product manufacturing process, it can also be the width of the interval from the start of manufacturing to the end of manufacturing.
[0043] The trigger condition refers to a condition for cutting out a monitoring target section to be the target of abnormality diagnosis from the monitoring target signal, and specifically indicates the cutting start time and cutting width of the monitoring target signal.
[0044] The trigger condition determination device 1 is implemented by a general-purpose information processing device such as a personal computer or a workstation, and includes an input unit 10 , an output unit 20 , a storage unit 30 , and a calculation unit 40 .
[0045] The input unit 10 is an input means for the operation unit 40, and is implemented by a data collection device, a keyboard, a pointing device, etc. In addition, the output unit 20 is implemented by a liquid crystal display, etc. The storage unit 30 is implemented by a hard disk device, etc. The storage unit 30 stores, for example, data (monitoring target signals, trigger candidate signals, trigger conditions, learning models, etc.) processed by the operation unit 40.
[0046] The operation unit 40 is implemented by, for example, a processor composed of a CPU (Central Processing Unit) and a memory (main storage unit) composed of a RAM (Random Access Memory), a ROM (Read Only Memory), etc. The operation unit 40 loads a program into a work area of the main storage unit to execute it, and controls each component by executing the program, thereby realizing a function that matches a predetermined purpose.
[0047] In addition, the computing unit 40 functions as a collecting unit (collecting means) 41, a cutting unit (cutting means) 42, a model generating unit (model generating means) 43, and a trigger condition determining unit (trigger condition determining means) 44 by executing the above-mentioned program. The details of each unit will be described later (refer to Figure 2 to Figure 8 ).
[0048] (Trigger condition determination method)
[0049] Reference Figure 2 to Figure 8The trigger condition determination method involved in this embodiment is described. The trigger condition determination method sequentially performs a collection process, a capture process, a model generation process, and a trigger condition determination process. In addition, in the trigger condition determination method, as described later, the capture process and the model generation process are repeatedly performed as needed.
[0050] <Collection process>
[0051] In the collection process, the collection unit 41 collects a signal group consisting of a monitoring target signal and a trigger candidate signal (step S1). In addition, here, the case where the collection unit 41 collects a plurality of monitoring target signals is described, but the collection unit 41 may collect only one monitoring target signal.
[0052] <Cutting process>
[0053] In the cutting process, the cutting unit 42 cuts out the monitoring target section of the monitoring target signal based on a predetermined criterion for the signal group collected in the collecting process. The cutting process will be described in detail below.
[0054] First, if Figure 3 As shown, the interception unit 42 performs a rough interception of the signal group (step S2). For example, in a device that performs repetitive actions such as a rolling device for coils, the monitoring object signal and the trigger candidate signal are continuously and uninterruptedly acquired for the continuously flowing coils. Therefore, in step S2, for example, the rough interception of the signal group is performed in order to divide the monitoring object signal and the trigger candidate signal for each coil. In addition, the rough interception of the signal group can be performed at a predetermined timing according to the type of the monitored device, or as shown in the figure, a rough interception signal can be selected from a plurality of trigger candidate signals and performed at the timing when the rough interception signal rises.
[0055] In addition, Figure 3 In the figure, reference numeral Sg represents the signal group before coarse interception, reference numeral Ss represents the monitoring object signal before coarse interception, reference numeral St represents the trigger candidate signal before coarse interception, reference numerals Sg1, Sg2, Sg3 represent the signal group after coarse interception, reference numerals Ss1, Ss2, Ss3 represent the monitoring object signal after coarse interception, and reference numerals St1, St2, St3 represent the trigger candidate signals after coarse interception.
[0056] Next, if Figure 4As shown in (a), the interception unit 42 selects a monitoring target signal (in the figure, the monitoring target signal Ss1) from the multiple monitoring target signals roughly intercepted in step S2, and intercepts the monitoring target interval Sm1 of the selected monitoring target signal Ss1 (step S3). The interception conditions when intercepting the monitoring target interval Sm1 in step S3 are determined based on the device characteristics of the monitoring target device. For example, when the monitoring target device is a "motor", the monitoring target signal Ss1 is a "motor current value", and it is determined whether the rising condition of the motor current value when the motor accelerates is normal, as shown in the figure, the interval of motor acceleration is set as the monitoring target interval Sm1. That is, the time when the motor starts to accelerate is designated as the interception start time of the monitoring target interval Sm1, and the interval from the start of motor acceleration to the end of acceleration is designated as the interception width.
[0057] Next, the cutting unit 42 calculates the correlation coefficient between the waveform included in the monitoring target interval Sm1 cut out in step S3 and the waveforms included in the other monitoring target signals Ss2 and Ss3 (step S4). Figure 4 As shown in (b), for other monitoring object signals Ss2 and Ss3, the interception unit 42 searches for the interval with the largest correlation coefficient with the waveform contained in the monitoring object interval Sm1 intercepted in step S3, thereby intercepting the monitoring object intervals Sm2 and Sm3 of the monitoring object signals Ss2 and Ss3 respectively (step S5).
[0058] Thus, in steps S4 and S5, a waveform similar to the waveform included in the monitoring target interval Sm1 extracted in step S3 is searched from the waveforms included in the monitoring target signals at other times roughly extracted in step S2. In addition, the method of searching for similar waveforms may use the Euclidean distance between the data of each time series signal, etc., in addition to the comparison of the correlation coefficients described above.
[0059] <Model creation process>
[0060] In the model generation process, for each signal group (multiple monitoring object signals), the start time of the monitoring object interval to be intercepted is determined in advance, and label data is generated to turn on (ON) the label at the start time and turn off (OFF) at other times. Through machine learning, a learning model is generated that takes the values of one or more trigger candidate signals at each time as input and the label data at each time as output.
[0061] In the model generation process, first, Figure 5As shown, for each signal group, the model generation unit 43 assigns a "trigger on (ON)" label to the values of the monitoring target signals Ss1, Ss2, Ss3 and the values of the trigger candidate signals St1, St2, St3 (hereinafter referred to as "the value of the signal group") corresponding to the start times p1, p2, p3 of the intercepted monitoring target intervals Sm1, Sm2, Sm3, and assigns a "trigger off (OFF)" label to the values of the signal group corresponding to the time other than the start time of the intercepted monitoring target intervals Sm1, Sm2, Sm3 (step S6). In addition, the "trigger on (ON)" label indicates that the value of the signal group assigned with the label is the interception start time, and the "trigger off (OFF)" label indicates that the value of the signal group assigned with the label is not the interception start time.
[0062] Next, the model generation unit 43 performs machine learning by taking the values of the signal group assigned the label “trigger on (ON)” and the values of the signal group assigned the label “trigger off (OFF)” as inputs and the labels “trigger on (ON)” and “trigger off (OFF)” as outputs, thereby Figure 6 As shown, a decision tree is generated (step S7).
[0063] Specifically, a decision tree can be generated using learning data that uses the target variable as the label of "trigger on (ON)" and "trigger off (OFF)" and the values of each trigger candidate signal corresponding to each moment of "trigger on (ON)" and "trigger off (OFF)" as explanatory variables. Here, "trigger on (ON)" can also be set to "1" and "trigger off (OFF)" can be set to "0" to be processed as a function. In addition, various machine learning models other than decision trees can also be used. Therefore, the learning model generated in step S7 is not limited to a decision tree, and can also be, for example, a random forest or a neural network. Here, "trigger on (ON)" can also be set to "1" and "trigger off (OFF)" can also be set to "0" to be processed as a function.
[0064] Here, in the model generation process, the trigger candidate signal included in the signal group is a single pulse signal, that is, Figure 7 The upper figure shows a signal of a single scan of the rise or fall of the ON signal in the ON-OFF signal. As shown in the lower figure of the figure, machine learning is performed after the trigger candidate signal is converted into a sawtooth wave.
[0065] like Figure 7As shown in the figure above, the single pulse signal is a signal that is turned on only for a short time. Therefore, in the above-mentioned interception process, when searching for a portion with a high similarity of the waveform, the single pulse signal should also be turned on when the "trigger is turned on", but sometimes the time of "trigger turning on" is offset before or after the time when the single pulse signal is turned on. On the other hand, as shown in the figure, by converting the single pulse signal into a sawtooth wave, the problem caused by the delay of turning on and off the single pulse signal can be eliminated.
[0066] The slope of the sawtooth wave after conversion is determined, for example, based on how many seconds you want to check after the single pulse signal is turned on, and is set to a slope that does not overlap with the rise of the next signal. Figure 7 As shown in part A of , it is preferred to have a margin (eg, about 5 scans) for a case where the interception of the monitoring target section starts before the single pulse signal is turned on by offsetting the interception start point.
[0067] In addition, if Figure 8 As shown in FIG. 1 , the shape of the sawtooth wave is defined by parameters tf and tb, but the relationship between the parameters tf and tb is preferably set to tb < tf. In addition, in the sawtooth wave, as shown in the figure, if the value z of the converted signal is within the range shown in B, it is determined that the trigger is on.
[0068] In the model generation process, when an error occurs during machine learning that the trigger condition cannot be generated normally, or when the discrimination accuracy cannot obtain the specified value, etc., the process returns to the above-mentioned interception process, shifts the monitoring object interval intercepted last time back and forth, and re-intercepts the monitoring object interval of the monitoring object signal, and then performs the model generation process again. That is, after re-intercepting the monitoring object interval of the monitoring object signal, the decision tree is constructed again. Moreover, when the trigger condition can be generated normally during machine learning, the model generation process is terminated, and when it cannot be generated normally, the process returns to the interception process again, and the interception process and the model generation process are re-performed.
[0069] like Figure 4 As shown, the trigger condition determination method involved in the present embodiment is a method of learning the state of the trigger candidate signal at the start time of the monitoring target interval (monitoring target interval Sm1) initially specified in the interception process. Therefore, for example, in the case where there is no feature in the state of the trigger candidate signal at the start time of the initially specified monitoring target interval, learning cannot be performed smoothly. Therefore, as described above, when an error occurs in the model generation process, the monitoring target interval initially specified in the interception process is shifted forward or backward, and the monitoring target interval is re-specified, thereby eliminating the bad condition during learning.
[0070] <Trigger condition determination process>
[0071] In the trigger condition determination process, the trigger condition determination unit 44 uses a decision tree to determine the trigger condition for the monitored signal for abnormal diagnosis (step S8). That is, the trigger condition determination unit 44 can know the timing of "trigger on (ON)" by inputting the trigger candidate signal into the decision tree, and can extract the timing of starting the monitoring of the monitored signal for abnormal diagnosis. Here, when the decision tree is used in step S8, the conditions of the generated decision tree are re-described to make it possible to concisely describe the trigger condition (refer to the following description). Fig.11 ). In this way, by using the decision tree generated in step S7, it is possible to easily understand under what conditions the monitored signal and the trigger candidate signal become triggered on. In addition, even when a learning model other than a decision tree is used, it is sufficient to sequentially input the trigger candidate signals at each moment as explanatory variables and use the timing of obtaining the output of "trigger on (ON)" as the timing of "trigger on (ON)".
[0072] (Abnormality diagnosis method for monitored equipment)
[0073] The abnormality diagnosis method for the monitored device extracts the signal of the monitored section from the monitored signal according to the trigger condition determined by the trigger condition determination method described above, stores the signal in the storage unit 30, and performs abnormality diagnosis of the monitored device based on the stored signal.
[0074] According to the trigger condition determination device, trigger condition determination method, and abnormality diagnosis method of the monitored object device involved in the present embodiment as described above, a learning model that has learned the conditions under which the monitored object signal and the trigger candidate signal become triggered on (ON) is used, so that the trigger condition for cutting off the monitored object section from the monitored object signal can be automatically determined. In addition, according to the trigger condition determination device, trigger condition determination method, and abnormality diagnosis method of the monitored object device involved in the present embodiment, the trigger condition for cutting off the monitored object section of the monitored object signal can be automatically determined, so there is no need to study and determine the trigger condition manually, and the prior preparation required for abnormality diagnosis of the monitored object device can be simplified.
[0075] Example
[0076] Reference Figures 9 to 12 An embodiment of the operation result prediction method according to the present invention will be described. In this embodiment, a sizing press machine is used as a monitoring target machine, and the present invention is applied to the speed performance of the main motor of the sizing press machine. Fig. 9 The monitoring target signal and the trigger candidate signal in this embodiment are shown.
[0077] exist Fig. 9 In the example, the monitored signal is the speed performance of the main motor, and the following five trigger candidate signals are selected.
[0078] v1: According to the stamping load is turned on (ON)
[0079] v2: Speed command value of the main motor
[0080] v3: Speed command value of width adjustment motor (lower side of driver)
[0081] v4: Open during stamping (ON)
[0082] v5: Total load of the punch force sensor
[0083] like Fig. 9 As shown, in order to observe the rise of the signal, the interval of 2500 to 2800 scan is first designated as the monitoring target interval Sm. Fig.10 , a graph is shown in which a portion having a maximum correlation coefficient is cut out from the waveform included in the monitoring target signal at other times and overlapped, Fig.11 The decision tree constructed in the model generation process is shown in FIG. Fig.12 2 shows a graph in which the monitoring target signal (actual speed performance of the main motor) is cut out and superimposed based on the trigger condition extracted from the decision tree in the cut-out condition determination step.
[0084] like Fig.11 As shown in FIG. 1 , the two conditions "v5<-5.5" and "13290≤v2<14370" are extracted from the decision tree constructed in the model generation process. Then, considering the case where the monitored signal is the speed performance of the main motor, the latter condition including the speed command value v2 of the main motor is selected, and "the speed command value v2 of the main motor = 13830" is determined as the trigger condition. Furthermore, the result of intercepting the monitored section of the monitored signal for abnormal diagnosis based on the trigger condition is as follows: Fig.12 As shown, the cut waveforms can be overlapped. Fig.12 , a waveform that is separated from other waveforms is included (see waveform C), but such a waveform may be determined as a candidate for abnormality through analysis using statistics such as Q statistics.
[0085] By determining the trigger conditions in this way, a normal model for abnormal diagnosis can be constructed by extracting normal data from the collected performance data, for example, and performing principal component analysis. In addition, during normal operation, statistics such as the Q statistic based on the normal model are obtained for the signal extracted under the above trigger conditions, so that abnormal diagnosis can be performed.
[0086] Furthermore, in this embodiment, the case where the monitored equipment is an ironmaking process, particularly a finishing press equipment in a hot rolling mill, is described, but the application scope of the present invention is not limited to this field, and it can be applied to production equipment in all manufacturing processes of petroleum-related products, chemicals, etc., experimental equipment in research institutions, etc.
[0087] The above specifically describes the trigger condition determination method of the time series signal, the abnormality diagnosis method of the monitored device, and the trigger condition determination device of the time series signal involved in the present invention through the methods and embodiments for implementing the invention, but the subject matter of the present invention is not limited to the above description, but must be interpreted broadly based on the scope of protection claimed in this application. In addition, technical solutions obtained by various changes and modifications based on these records are of course also included in the subject matter of the present invention.
[0088] Description of Reference Numerals
[0089] 1…trigger condition determination device; 10…input unit; 20…output unit; 30…storage unit; 40…calculation unit; 41…collection unit (collection unit); 42…interception unit (interception unit); 43…model generation unit (model generation unit); 44…trigger condition determination unit (trigger condition determination unit).
Claims
1. A method for determining a trigger condition of a time series signal, wherein, when performing abnormality diagnosis of a monitored device, a condition, i.e., a trigger condition, is determined for cutting out a monitored section, i.e., a monitored signal, i.e., a time series signal representing the state of the monitored device, which is a monitored section, to be the subject of the abnormality diagnosis, wherein: include: A collecting step, in which a signal group consisting of one or more monitoring target signals related to the monitoring target device and a trigger candidate signal is collected, wherein the trigger candidate signal represents a time series signal that can become the trigger condition among time series signals related to the monitoring target device and detected at the same time as the monitoring target signal; a cutting step of cutting out a monitoring target section of the monitoring target signal based on a predetermined reference for the signal group; A model generation step, in which, for the signal group, the start time of the intercepted monitoring target interval is determined, label data is generated to turn on the label at the start time and turn off the label at other times, and a learning model is generated by machine learning, which takes one or more of the trigger candidate signals at each time as input and takes the label data at each time as output; as well as A trigger condition determination step is performed in which the trigger condition is determined using the learning model for the monitoring target signal for performing the abnormality diagnosis.
2. The trigger condition determination method of a time series signal according to claim 1, characterized in that: The cutting step cuts out a monitoring target section based on a device characteristic of the monitoring target device for a first monitoring target signal selected from the plurality of monitoring target signals collected in the collecting step, The cutting step searches for a section having a maximum correlation coefficient with a waveform included in a monitoring target section of the first monitoring target signal among the plurality of monitoring target signals other than the first monitoring target signal, thereby cutting out each monitoring target section.
3. The method for determining the trigger condition of a time series signal according to claim 1 or 2, characterized in that: The learning model is a decision tree.
4. The method for determining the trigger condition of a time series signal according to claim 1 or 2, characterized in that: In the model generation step, when the candidate trigger signal is a single pulse signal, machine learning is performed after the candidate trigger signal is converted into a sawtooth wave.
5. The method for determining the trigger condition of a time series signal according to claim 1 or 2, characterized in that: When the discrimination accuracy cannot reach the specified value during machine learning in the model generation process, the process returns to the interception process, shifts the monitoring object interval intercepted last time forward and backward, and after re-intercepting the monitoring object interval of the monitoring object signal, the model generation process is performed again.
6. A method for diagnosing abnormality of a monitored device, characterized in that: According to the trigger condition determined by the trigger condition determination method of the time series signal described in any one of claims 1 to 5, the signal of the monitoring object section that becomes the object of the abnormal diagnosis is intercepted and accumulated from the time series signal representing the state of the monitored object device, that is, the monitored object signal, and the abnormal diagnosis of the monitored object device is performed based on the accumulated signal.
7. A time series signal trigger condition determination device, which, when performing abnormality diagnosis of a monitored object device, determines a condition, i.e., a trigger condition, for cutting out a monitored object section, i.e., a monitored object signal, i.e., a time series signal representing the state of the monitored object device, which is a monitored object signal, to be the subject of the abnormality diagnosis, characterized in that: include: a collecting unit for collecting a signal group consisting of one or more monitoring target signals related to the monitoring target device and a trigger candidate signal, wherein the trigger candidate signal represents a time series signal that can become the trigger condition among time series signals related to the monitoring target device and detected at the same time as the monitoring target signal; an interception unit, for intercepting a monitoring target section of the monitoring target signal based on a predetermined reference for the signal group; A model generation unit, for the signal group, determines the start time of the intercepted monitoring target interval, generates label data that turns on the label at the start time and turns off the label at other times, and generates a learning model through machine learning that takes one or more of the trigger candidate signals at each time as input and takes the label data at each time as output; as well as The trigger condition determination unit determines the trigger condition using the learning model for the monitoring target signal for performing the abnormality diagnosis.
