Control method, control device, mechanical equipment and recording medium

Abstract

The present application relates to a control method, a control device, a mechanical equipment and a recording medium. The control device includes a controller. The controller is configured to obtain a measurement value of a state of a mechanical equipment corresponding to a period of time when the mechanical equipment reaches a second state from a first state, extract at least one predetermined feature value by using the measurement value, and extract data for machine learning from the data of the at least one predetermined feature value based on a separation degree for distinguishing the first state from the second state.

Description

Technical Field

[0001] The present invention relates to a control method, a control device, a mechanical equipment including the control device, a control program, and a computer-readable recording medium. Background Art

[0002] The operating condition of mechanical equipment may change at any time depending on changes in the conditions of constituent parts, etc. If an operating condition within an allowable range based on the purpose of use of the mechanical equipment is called a normal state, and an operating condition outside the allowable range is called a fault state, then, for example, in the case where a manufacturing machine is in a fault state, a fault such as the manufacture of a defective product or a stoppage of a production line occurs.

[0003] In the case of a manufacturing machine, etc., even if the same operation is repeatedly and continuously performed, maintenance operations are generally performed regularly or irregularly to suppress the occurrence of a fault state as much as possible. Although it is effective to shorten the execution interval between maintenance operations in order to improve preventive safety, since the manufacturing machine, etc. is stopped during the maintenance operation, if the frequency of maintenance operations is excessively increased, the operating rate of the manufacturing machine, etc. is reduced. Therefore, it is expected that such a state can be detected when the occurrence of a fault state is approaching while the machine, etc. is still in a normal state. This is because, in this case, when the approach of a fault state is detected (that is, when the occurrence of a fault is predicted), maintenance operations of the machine, etc. can be performed, so that an excessive reduction in the operating rate can be suppressed.

[0004] As a method for predicting the occurrence of a fault, a method is known in which a learned model generated by machine learning the state of a mechanical device is prepared in advance and the state of the mechanical device is evaluated at an evaluation time by using the learned model. In order to improve the prediction accuracy, it is important to construct a learned model suitable for fault prediction. However, for this purpose, it is important whether the learning data used for machine learning is suitable.

[0005] For example, Japanese Patent Laid-Open No. 2011-70635 discloses extracting a vector based on a sensor signal indicating a state of mechanical equipment and selecting features to be used based on data inspection of feature vectors. In addition, Japanese Patent Laid-Open No. 2011-70635 discloses generating a normal model of mechanical equipment based on selected learning data.

[0006] In addition, regarding selection from multiple learning data prepared for the corresponding season according to seasonal changes, Japanese Patent Laid-Open No. 2011-59790 discloses selecting sensor signals that should be paid attention to and selected according to the abnormality based on abnormality measurements and evaluation results of the influence of each sensor signal as a result of multivariate analysis.

[0007] In mechanical equipment, measurement data for various parameters are obtained to manage the operating conditions of the mechanical equipment. However, since the frequency of occurrence of failures is generally not high, it is not easy to collect data on a large number of failure conditions. At the same time, the huge amount of measurement data measured for various parameters includes both data related to the occurrence of failures and data unrelated to the occurrence of failures, and therefore is redundant as learning data for machine learning.

[0008] Although selection of learning data is recognized in Japanese Patent Laid-Open No. 2011-70635 and Japanese Patent Laid-Open No. 2011-59790, there is no sufficient discussion on how to actually select and obtain learning data.

[0009] In particular, in the case of mechanical equipment where the frequency of failure occurrence is low and from which data of only a small amount of failure cases can be collected, it is difficult to generate a learned model with high prediction accuracy by conventional methods.

[0010] Therefore, a method is desired that can generate a failure prediction model with high prediction accuracy even when only a small amount of data on failure conditions can be collected from mechanical equipment. Summary of the invention

[0011] According to a first aspect of the invention, a control method includes obtaining a measurement value of a state of a mechanical equipment corresponding to a time period in which the mechanical equipment reaches a second state from a first state, extracting at least one predetermined characteristic value by using the measurement value, and extracting data for machine learning from the data of the at least one predetermined characteristic value based on a degree of separation for distinguishing the first state from the second state from each other.

[0012] According to a second aspect of the present invention, a control device includes a controller. The controller is configured to obtain a measurement value of a state of a mechanical device corresponding to a period of time when the mechanical device reaches a second state from a first state, extract at least one predetermined feature value by using the measurement value, and extract data for machine learning from the data of the at least one predetermined feature value based on a separation degree for distinguishing the first state from the second state from each other.

[0013] Further features of the present invention will become apparent from the following description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a schematic functional block diagram for describing functional blocks included in the failure prediction system of the embodiment.

[0015] Figure 2 is a schematic diagram for describing the hardware configuration of the embodiment.

[0016] Figure 3is a schematic diagram for describing a method for extracting a feature value according to an embodiment.

[0017] Figure 4 is a schematic diagram for describing a method for calculating a degree of separation according to an embodiment.

[0018] Figure 5 is a schematic diagram for describing a machine learning method according to an embodiment.

[0019] Figure 6 is a schematic diagram for describing a fault prediction method according to an embodiment.

[0020] Figure 7 is a flowchart illustrating a process for generating a fault prediction model according to an embodiment.

[0021] Figure 8 is a flowchart illustrating a processing procedure of failure prediction according to an embodiment.

[0022] Fig. 9 is a diagram illustrating an example of an operation assistance screen displayed in the embodiment.

[0023] Fig.10 is a diagram illustrating another example of the operation assistance screen displayed in the embodiment.

[0024] Fig.11 is a schematic diagram for describing a method for determining a determination threshold according to an embodiment. DETAILED DESCRIPTION

[0025] As an embodiment of the present invention, a fault prediction system, a control method, a control device, a mechanical equipment including the control device, a control program, a computer-readable recording medium, etc. for predicting a fault of a mechanical equipment will be described with reference to the accompanying drawings. In the following description, considering the purpose of use of the mechanical equipment, a state in which the operating condition is within the allowable range may be referred to as a normal state, and a state in which the operating condition is outside the allowable range may be referred to as a fault state or an abnormal state.

[0026] Configuration of function blocks

[0027] Figure 1 is a schematic functional block diagram for describing the configuration of functional blocks included in the fault prediction system of the embodiment. Figure 1 The functional elements required for describing the features of this embodiment are indicated as functional blocks, but illustrations of general functional elements that are not directly related to the problem-solving principle of the present invention are omitted. Figure 1Each functional element shown in the figure is functionally conceptual and is not necessarily physically configured as shown. For example, the specific configuration of the distribution and integration of each functional block is not limited to the illustrated example, and all or part of it can be functionally or physically distributed or integrated in any unit according to the conditions of use, etc.

[0028] like Figure 1 As shown in , the failure prediction system of this embodiment includes a mechanical equipment 10 as a diagnosis target and a failure prediction device 100 .

[0029] The mechanical equipment 10 is a variety of industrial equipment that manufactures products as result items by assembling predetermined workpieces (such as a multi-joint robot or a manufacturing device installed in a production line). The mechanical equipment 10 includes various sensors 11 for measuring the state of the mechanical equipment 10. For example, in the case where the mechanical equipment 10 is a multi-joint robot, a sensor for measuring the current value of the motor driving the joint, an angle sensor of the joint, a sensor for measuring speed, vibration or sound, etc. may be provided. However, these are merely examples, and appropriate types and quantities of sensors may be provided as the sensors 11 at appropriate locations depending on the type of the mechanical equipment 10, the purpose of use, etc. As the sensor 11, various sensors such as force sensors, torque sensors, vibration sensors, sound sensors, image sensors, distance sensors, temperature sensors, humidity sensors, flow rate sensors, pH sensors, pressure sensors, viscosity sensors, and gas sensors may be used. It should be noted that although for ease of explanation in Figure 1 Only a single sensor 11 is shown, but typically a plurality of sensors would be provided.

[0030] The mechanical equipment 10 is communicably connected to the failure prediction device 100 in a wired or wireless manner, and the failure prediction device 100 can obtain data measured by the sensor 11 through communication.

[0031] In the generation stage of the fault prediction model, the fault prediction device 100 selects a feature value highly correlated with the occurrence of a fault of the mechanical equipment 10 by using the data collected from the sensor 11, and generates and stores a learned model (i.e., a fault prediction model) by machine learning using the selected feature value. In addition, in the evaluation stage (i.e., the fault prediction stage), the data of the evaluation time collected from the sensor 11 is input into the learned model, the deviation is calculated by using the input and output of the learned model, and it is determined whether the occurrence of a fault is imminent. In the following, the functional blocks included in the fault prediction device 100 will be described sequentially.

[0032] The fault prediction device 100 includes a controller 110 , a storage portion 120 , a feature value selection portion 131 , a display portion 130 , and an input portion 140 .

[0033] The controller 110 and the feature value selection part 131 include a plurality of functional blocks, and these functional blocks are constituted by, for example, a central processing unit CPU of the fault prediction apparatus 100 reading and executing a control program stored in a storage device. Alternatively, part or all of the functional blocks may be constituted by hardware included in the fault prediction apparatus 100 (such as an application-specific integrated circuit ASIC).

[0034] The storage section 120 includes a sensor data storage section 121, a feature value storage section 122, a separation degree storage section 123, a fault prediction model condition storage section 124, and a fault prediction model storage section 125. Each of these sections included in the storage section 120 is configured by appropriately assigning to a storage area of ​​a storage device such as a hard disk drive, a random access memory RAM, or a read-only memory ROM. The storage section 120 is a data acquisition section that acquires and stores various data required for processing to perform fault prediction.

[0035] The display portion 130 and the input portion 140 are user interfaces included in the fault prediction apparatus 100. A display device such as a liquid crystal display or an organic electroluminescent display is used as the display portion 130, and an input device such as a keyboard, a jog dial, a mouse, a pointing device, or a sound input device is used as the input portion 140.

[0036] The sensor data collecting part 111 of the controller 110 obtains measurement data from the sensor 11 of the mechanical equipment 10 and stores the measurement data in the sensor data storing part 121. That is, for example, measurement data on the state of the mechanical equipment 10, such as the temperature, current, speed, pressure, vibration, sound, etc. of each part measured in the mechanical equipment 10, is collected and stored.

[0037] The feature value extraction part 112 extracts a feature value indicating the characteristics of the operating condition (operating condition) of the mechanical equipment 10 based on the measurement data stored in the sensor data storage part 121, and stores the feature value in the feature value storage part 122. For example, as feature value data, the maximum value and / or minimum value of the measurement value of the sensor 11 collected in one operation cycle of the mechanical equipment 10 can be extracted, or the average value of the measurement value can be calculated. Alternatively, the value of the sensor in a predetermined period can be converted into a time series frequency domain by integration. In addition, the differential value or the secondary differential value of the sensor value arranged in a time series with respect to time can be used as feature value data. In addition, in the case where the measurement value of the sensor 11 itself (that is, the raw data of the measurement value) is useful as a determining factor for detecting whether a fault is imminent, the measurement value itself can be used as feature value data. In the present embodiment, the feature value extraction part 112 extracts or calculates a feature value based on the measurement value of the sensor 11, and generates time series feature value data and stores it in the feature value storage part 122.

[0038] The separation degree calculation part 113 obtains the characteristic value from the characteristic value storage part 122 and calculates the maximum value of the separation degree for each characteristic value, and stores the maximum value in the separation degree storage part 123. The separation degree is an indicator indicating whether the characteristic value of the mechanical equipment 10 is sensitive to the state change when the mechanical equipment 10 transitions from the normal state to the fault state, and the details will be described later.

[0039] The feature value selection part 131 obtains the maximum value of the degree of separation of each feature value from the degree of separation storage part 123, and selects the feature value to be used for machine learning based on the obtained information and the threshold value stored in the degree of separation threshold setting part 132. In addition, the normal interval setting part 133 sets the normal interval of the mechanical equipment 10 with reference to the timing at which the maximum value of the degree of separation appears for the selected feature value. The normal interval is a period of time when the mechanical equipment 10 is in a normal state (where no signs of failure have yet appeared in the feature value). The selected feature value and the normal interval are stored in the fault prediction model condition storage part 124 as conditions for extracting learning data for machine learning.

[0040] When generating the learned model, the data extraction part 114 extracts data for machine learning from the feature value storage part 122 based on the conditions stored in the fault prediction model condition storage part 124, and outputs the data for machine learning to the fault prediction model generation part 115. In addition, at the evaluation time, the data extraction part 114 extracts feature values ​​of the same kind as the feature values ​​stored in the fault prediction model condition storage part 124 (i.e., feature values ​​for machine learning) from the feature value data corresponding to the evaluation time and stored in the feature value storage part 122, and outputs the extracted feature values ​​to the fault determination part 116.

[0041] When generating a learned model, the fault prediction model generating part 115 generates a learned model (ie, a fault prediction model) by using the data for machine learning input from the data extracting part 114 , and stores the learned model in the fault prediction model storing part 125 .

[0042] At the evaluation time, the fault determination section 116 inputs the feature value data for evaluation input from the data extraction section 114 to the learned model (i.e., the fault prediction model stored in the fault prediction model storage section 125), and calculates the degree of deviation between its input and output. Then, the fault determination section 116 determines whether there is a sign of a fault by comparing the degree of deviation with a determination threshold.

[0043] The failure notification section 117 notifies an external device of the determination result of the failure determination section 116 , or displays the determination result on the display section 130 .

[0044] Hardware Configuration

[0045] Figure 2 An example of the hardware configuration of the fault prediction system of this embodiment is schematically illustrated. Figure 2 As shown in , the fault prediction system may include personal computer hardware, which includes a CPU 1601 as a main controller, and a ROM 1602 and a RAM 1603 as a storage part. ROM 1602 can store information such as processing programs and reasoning algorithms for implementing the fault prediction method to be described later. In addition, when the control process of the method is executed, RAM 1603 is used as a work area of ​​CPU 1601, etc. In addition, an external storage device 1606 is connected to the control system of the fault prediction system. The external storage device 1606 is composed of a hard disk drive HDD, a solid-state device SSD, an external storage part of another system installed thereon via a network, etc.

[0046] The control program for the CPU 1601 to implement the fault prediction method of the present embodiment, which will be described later, can be stored in a storage portion such as an external storage device 1606 constituted by an HDD or SSD, or in an electrically erasable programmable ROM area (EEPROM area) such as a ROM 1602. In this case, the processing program for the CPU 1601 to implement the fault prediction method is supplied to each of the above-mentioned storage portions through the network interface NIF 1607, and can be updated to a new program (i.e., a different program). Alternatively, the processing program for the CPU 1601 to implement the fault prediction method can be supplied to each of the above-mentioned storage portions via various storage media such as magnetic disks, optical disks, and flash memories, and drive devices for these storage media, and its content can be updated. Various storage media, storage portions, or storage devices storing programs that the CPU 1601 can use to execute processing for implementing the fault prediction method constitute a computer-readable recording medium that stores the fault prediction process of the present invention.

[0047] CPU 1601 is connected to Figure 1 The sensor 11 shown in FIG. Figure 2 1 and 16. For a simpler illustration, the sensor 11 is shown as being directly connected to the CPU 1601, but the sensor 11 may be connected to the CPU 1601 via, for example, IEEE 488 (the so-called general purpose interface bus GPIB). In addition, the sensor 11 may be connected to the CPU 1601 via a network interface 1607 and a network 1608.

[0048] The network interface 1607 may be configured by, for example, using a communication standard for wired communication (such as IEEE 802.3) or a communication standard for wireless communication (such as IEEE 802.11 or 802.15). The CPU 1601 may communicate with other devices 1104 and 1121 via the network interface 1607. For example, in the case where a robot is used as a target for fault prediction, the devices 1104 and 1121 may be an integral control device, a management server, or the like, such as a programmable logic control PLC or a sequencer, deployed to control or manage the robot.

[0049] exist Figure 2 In the example shown, Figure 1The operation part 1604 and the display device 1605 related to the input part 140 and the display part 130 shown in FIG. 1 are connected to the CPU 1601 as a user interface device (UI device). The operation part 1604 can be composed of a terminal such as a handheld terminal, a device such as a keyboard, a jog dial, a mouse, a pointing device, a sound input device, or a control terminal including these. The display device 1605 can be any device as long as it can display information related to the processing performed by the separation degree calculation part 113, the fault prediction model generation part 115, the fault determination part 116, etc. on its display screen, and a liquid crystal display device can be used, for example.

[0050] Fault prediction methods

[0051] In the present embodiment, the fault prediction model generating part 115 of the fault prediction device 100 constructs a learned model used as a fault prediction model by so-called unsupervised learning. In order to learn the fault characteristics of mechanical equipment by unsupervised learning, machine learning is performed by using only the operating data of the state without faults (that is, the operating data of the period when the mechanical equipment is operating normally). In unsupervised learning, the distribution of input data is learned by providing only a large amount of input data to the learning device. That is, unsupervised learning is a method of making a device that performs processing such as compression, classification, and deformation on input data learn processing without providing the device with teacher output data corresponding to the input data.

[0052] A fault prediction method using an unsupervised learning method will be described in detail. Machine learning is performed by setting the operating condition within the allowable range as a normal state in consideration of the purpose of use of the mechanical equipment and using only the operating data of the mechanical equipment in the normal state. In this embodiment, an autoencoder is used as an unsupervised learning model.

[0053] This embodiment is characterized by a method for extracting data for machine learning, and in this embodiment, characteristic values ​​that are sensitive to state changes when mechanical equipment transitions from a normal state to a fault state are selected, and data of the selected characteristic values ​​in a predetermined time period used as a normal interval are used as learning data.

[0054] First, refer to Figure 3 Describe the extraction of feature values. Figure 3As shown in , it is assumed that the mechanical equipment 10 used as the fault prediction target includes a sensor 11, and the sensor 11 includes a sensor 1 as a current sensor, a sensor 2 as a speed sensor, and a sensor 3 as a pressure sensor. In the fault prediction method according to the present embodiment, first, a feature value indicating the operating condition (operating condition) of the mechanical equipment 10 is extracted based on the measurement data of each sensor included in the sensor 11. The feature value is extracted by performing an integral transformation of the time series measurement data of the sensor 11 to the frequency domain, calculating the first differential or second differential of the measurement data with respect to time, performing a filtering process on the measurement data, and extracting the maximum value and the minimum value of the periodic operation from the measurement data. It should be noted that the above-mentioned sensors and the processing of the measurement data are only examples, and any sensor and the processing of the measurement data can be adopted as long as data suitable for grasping the state of the mechanical equipment can be obtained. In addition, if the state of the mechanical equipment can be easily analyzed by using the measurement data of the sensor itself, the measurement data itself can be used as a feature value without performing any special processing.

[0055] Figure 3 The state in which time series data of nine kinds of feature values ​​shown on the right are extracted by performing three kinds of processing on each measurement data of sensor 1 , sensor 2 , and sensor 3 is schematically illustrated.

[0056] Next, a method for selecting a feature value that sensitively changes when mechanical equipment transitions from a normal state to a fault state from among the extracted feature values ​​will be described.

[0057] First, the controller 110 extracts the time series data of the feature value corresponding to the period when the mechanical equipment 10 reaches the fault state from the normal state, that is, Figure 3 Next, the separation degree is calculated for each of the nine characteristic values, and the separation degree is an indicator of the sensitivity of the mechanical equipment 10 to transition from a normal state to a fault state.

[0058] Will refer to Figure 4 A method for calculating the degree of separation based on time series eigenvalue data is described. Here, the method for calculating the degree of separation is described by taking Fisher's linear discriminant analysis as an example. Fisher's linear discriminant analysis is a method for determining a straight line that best distinguishes two sets from each other. In this embodiment, the two sets are a normal state and a fault state. It can be determined which set a eigenvalue belongs to by checking which side of an arbitrary point on the time axis the eigenvalue is located on.

[0059] Specifically, when the data set on the left side of an arbitrary point of the time series feature value is defined as A and the data set on the right side of the arbitrary point is defined as data set B, the mean value μ and the variance σ are obtained for each of the data sets A and B. Using the obtained mean value μ and the obtained variance σ of each set, the separation degree is obtained according to the following formula.

[0060] [Formula 1]

[0061]

[0062] While moving the boundary between data sets A and B along the timeline, the above process is performed for each of the 9 eigenvalues, thereby obtaining the separation degree of each point in the timeline. Then, for each of the 9 eigenvalues, the maximum value among the separation degrees obtained in time series is obtained. The eigenvalue whose maximum value of separation degree is large can be considered as a eigenvalue having higher distinguishability between normal state and fault state (i.e., higher sensitivity to fault occurrence). In addition, the timing at which the maximum separation degree is obtained for the eigenvalue can be regarded as the timing at which the normal state and the fault state can be best distinguished for the eigenvalue.

[0063] Fig. 9 An example of an operation assistance screen displayed on the display section 130 of the fault prediction device 100 when the operator performs an operation of selecting learning data and generating a learned model through machine learning is illustrated.

[0064] A field for displaying information for specifying a fault condition is disposed in the upper left portion of the screen. This field is provided to facilitate the operator to perform operations such as specifying, changing, and checking which fault condition's measurement data will be used as learning data.

[0065] A field for displaying the range of the maximum value of the separation degree is disposed in the upper right portion of the screen. This field is provided to facilitate the operator to perform operations such as specifying, changing, and checking the range of the maximum value of the separation degree, which is a condition for selecting the feature value to be used for learning data. Fig. 9 An example is illustrated in which the maximum value of the degree of separation is 50 or more used as a condition for selecting feature values ​​for learning data.

[0066] In the middle portion of the screen, information related to the feature values ​​selected as matching the selection condition is arranged so that a feature value having a greater maximum value of separation is displayed at a higher position. Fig. 9The diagram shows an example in which the maximum value of the current, the minimum value of the pressure, and the average value of the speed have matched the selection conditions among the 9 characteristic values, and the maximum value of the separation degree is greater in this order. For example, regarding the maximum value of the current, it is indicated that the maximum value of the separation degree is 98, which has been detected 7 days before the fault occurs. It should be noted that although Fig. 9 In the example, feature values ​​extracted from measurement data of different kinds of sensors (such as current, speed, and pressure) are selected, but this is just a coincidence, and, for example, feature values ​​obtained by performing different processing on the measurement values ​​of the same sensor may also be selected.

[0067] The graph fields displayed on the right side of the center portion of the screen each include a graphical representation of a graph of changes over time in a characteristic value, a timing at which a separation value of the characteristic value reaches a maximum value, and a normal interval which will be described below.

[0068] In this embodiment, the data of the period when the mechanical equipment is in a normal state (i.e., the data of the normal interval among the time series data of the selected characteristic value) is used as the learning data, and the normal interval designation information field is displayed at the lower part of the screen. This field is provided to facilitate the operator to instruct setting, checking and changing the normal interval. Fig. 9 An example is illustrated in which a period of 10 days or more before the occurrence of a failure is designated as a normal interval.

[0069] The normal interval is set to a past interval earlier than any timing at which the maximum value of the separation degree is detected for the selected three feature values. This is because if the timing at which the maximum value of the separation degree is detected for the selected feature value is within the normal interval, the data of the feature value within the normal interval includes data of the fault state, which is not convenient for accurate machine learning of the characteristics of the normal state. Fig. 9 In the example of , among the selected feature values, the timing at which the maximum value of the separation has been detected is the earliest for the speed average value, but there is no problem because the time period at which the separation reaches the maximum value earlier than the speed average value is set as the normal interval.

[0070] It is to be noted that when a period earlier than the time point at which the fault occurs (i.e., a past period farther from the time point at which the fault occurs) is set as a normal interval, data of a normal state particularly close to the initial state (i.e., data of a state in which the mechanical equipment is far from the fault) can be used for machine learning. It is to be noted that when a period earlier than the time point at which the fault occurs (i.e., a past period farther from the time point at which the fault occurs) is set as a normal interval, the normal interval becomes shorter in time, and thus the amount of data that can be used for machine learning becomes smaller. Therefore, it is desirable to set the normal interval by comprehensively considering the quality and quantity of the data.

[0071] The selection field displayed at the left end of the middle portion of the screen is provided to facilitate the operator to perform operations such as checking whether a feature value is selected, instructing selection / non-selection of a feature value, and changing selection / non-selection of a feature value. For example, as a result of specifying a selection condition of a separation degree of 50 or more, even in the case where the above three feature values ​​are automatically selected, if the operator desires to cancel the selection of a part of the three feature values, the selection can be canceled by unchecking the selection field. Fig.10 An example of a display screen in a case where the operator has instructed to exclude the average value of speed from the feature values ​​used for machine learning is illustrated.

[0072] As described above, according to the present embodiment, feature values ​​having high sensitivity for distinguishing a normal state from a faulty state can be selected, and for the selected feature values, data of a normal period in which mechanical equipment is in a normal state can be extracted as machine learning data.

[0073] The model generation button disposed at the right end of the lower portion of the screen is an icon for causing the operator to instruct the fault prediction device to start an operation of generating a fault prediction model through machine learning using the selected feature values.

[0074] Next, we will refer to Figure 5 (This figure is a schematic diagram) A method for causing an autoencoder to perform machine learning by using the learning data extracted as described above is described. An autoencoder is a neural network that compresses (i.e., encodes) input learning data into data of a smaller data amount and then restores (i.e., decodes) the data. The autoencoder learns "parameters for appropriately compressing and restoring input data", i.e., the characteristics of the input data.

[0075] The autoencoder encodes the input value x to compress the input value x into the intermediate layer z. Then, the autoencoder decodes the intermediate layer z to restore the intermediate layer z to the output value y. The autoencoder performs machine learning so that the restored difference J between the input value and the output value becomes smaller.

[0076] That is, the auto encoder determines W and b of the following Formula 2 and W' and b' of the following Formula 3 so that the restoration difference J in the following Formula 4 becomes small. Note that s represents an activation function.

[0077] [Formula 2]

[0078] z=s(Wx+b)

[0079] [Formula 3]

[0080] y=s(W′x+b′)

[0081] [Formula 4]

[0082] Restoration difference J = Σ(xy) 2

[0083] When data having characteristics similar to those of the learning data is input, the autoencoder that has performed learning performs encoding and decoding by using the parameters obtained by the learning to output an output value with a small recovery difference. The autoencoder that has performed learning will sometimes be described as a post-learning model or a fault prediction model. On the contrary, when data having characteristics different from those of the learning data is input to the post-learning model, encoding and decoding cannot be successfully performed by using the parameters obtained by the learning, and therefore the recovery difference is large.

[0084] In this embodiment, in order to utilize this property, machine learning is performed by inputting data of a normal interval of the selected feature value as an input value x to the autoencoder.

[0085] In addition, when fault prediction is performed, data at the evaluation time of the same kind of feature value as that selected during learning is input as input value x to the post-learning model to output output value y. Then, the recovery difference between the input value x and the output value y is calculated, and the recovery difference (i.e., the degree of deviation between the input and the output) is used as an indicator indicating the degree of deviation of the mechanical equipment from the normal state.

[0086] In addition, in the present embodiment, a determination threshold is pre-set for determining whether the occurrence of a failure of the mechanical equipment is imminent by using the degree of deviation. In order to set the determination threshold, first, characteristic value data corresponding to the time period from the normal state to the occurrence of the failure based on the sensor data of the actual mechanical equipment is input, and the time change of the degree of deviation until the failure occurs is studied. As this characteristic value data, data of an extracted characteristic value of the same kind as the characteristic value selected during learning (that is, a characteristic value obtained by performing the same processing on the measurement data of the same sensor) is used. The determination threshold for determining that the occurrence of a failure is imminent is set based on the time change of the degree of deviation. In the case where the degree of deviation is equal to or greater than the determination threshold, it is determined that the occurrence of a failure of the mechanical equipment is imminent, that is, there is a sign of a failure.

[0087] Fig.11 is a diagram for describing in detail a method for determining a threshold value. Fig.11In the graph of , the horizontal axis represents time, the vertical axis represents an index value, the index value indicates the proximity of the occurrence of a fault (i.e., the degree of deviation between the input and output of the model after learning), and the graph indicates the change in the index value over time from the initial stage of the normal state to the occurrence of the fault (i.e., the abnormal state). Assume that it is desired to ensure the situation at a predetermined time t before the occurrence of the fault after the fault prediction device has predicted and notified the proximity of the occurrence of the fault, that is, it is expected that the fault prediction device predicts the occurrence of the fault at a time t earlier than the occurrence of the fault. In this case, the index value at a time t earlier than the occurrence of the fault (i.e., the degree of deviation between the input and output of the model after learning) as shown in the figure is set as the determination threshold T for fault prediction. This serves as a determination threshold setting step.

[0088] Fault prediction using the above-described learned model and determination threshold value will be described. Figure 6 is a schematic diagram for describing a fault prediction method using an autoencoder.

[0089] Evaluation data indicating the operating state of the mechanical equipment at the evaluation time is input to the post-learning model, and the degree of deviation indicating how much the state of the mechanical equipment differs from the learned normal state is calculated by using the input value and the output value. As the evaluation data, data at the evaluation time for the same kind of feature value as that selected during learning (i.e., feature value obtained by performing the same processing on the measurement data of the same sensor) is used. Specifically, as Figure 6 As shown in , the evaluation data is input into the fault prediction model, and the recovery difference J between the output value y of the fault prediction model obtained as the input result and the input value x is calculated, and used as the deviation from the normal state. In the present embodiment, this deviation is used as an index value indicating the proximity of the occurrence of the fault. In the case where the deviation (i.e., the recovery difference J) is equal to or greater than the determination threshold T, it is determined that the time until the fault occurs is equal to or less than the predetermined time t, that is, it is determined that there is a sign of a fault. On the contrary, in the case where the deviation (i.e., the recovery difference J) is less than the determination threshold T, it is determined that the time until the fault occurs is longer than the predetermined time t, that is, it is determined that there is no sign of a fault.

[0090] Process

[0091] Next, we will refer to Figure 7 and Figure 8 The flowchart of FIG. 1 describes the procedure of the processing performed by the failure prediction device 100 .

[0092] Figure 7 is a flow chart illustrating a process for generating a fault prediction model.

[0093] First, in step S101, the sensor data collecting section 111 of the failure prediction apparatus 100 obtains measurement data from the sensor 11 for measuring the state of the mechanical equipment 10, and stores the measurement data in the sensor data storing section 121. This serves as a measurement data obtaining step.

[0094] Next, in step S102, the feature value extraction section 112 extracts feature values ​​indicating characteristics of the operating condition of the mechanical equipment 10 based on the sensor data stored in the sensor data storage section 121, and stores the extracted feature values ​​in the feature value storage section 122. This serves as a feature value extraction step.

[0095] Next, in step S103, the controller 110 determines whether a fault has occurred in the mechanical equipment 10. The occurrence of a fault may be input to the controller 110 of the fault prediction device 100 by a user via the input section 140, or input to the controller 110 of the fault prediction device 100 by a control section (not shown) of the mechanical equipment 10 via the network interface NIF 1607.

[0096] In the case where no failure occurs, that is, in the case where the result of step S103 is NO, the process returns to step S101, and steps S101 and S102 are repeated until a failure occurs.

[0097] In the case where a failure has occurred, that is, in the case where the result of step S103 is YES, the process proceeds to step S104, and the separation degree calculation section 113 obtains feature values ​​from the feature value storage section 122 and calculates the separation degree for each feature value.

[0098] Next, in step S105, the separation degree calculation section 113 extracts the maximum value of the separation degree from the time series data of the separation degree calculated for each characteristic value, and stores the maximum value in the separation degree storage section 123. That is, for each of the plurality of characteristic values, the maximum value of the separation degree in the period from the normal state to the fault state is obtained. This serves as a separation degree maximum value obtaining step.

[0099] Next, in step S106, the feature value selection section 131 obtains the maximum value of the degree of separation for each feature value from the degree of separation storage section 123, and selects the feature value to be used for machine learning based on the threshold value stored in the degree of separation threshold setting section 132. In addition, the normal interval setting section 133 sets the normal interval of the mechanical equipment with reference to the timing of reaching the maximum value of the degree of separation of the selected feature value, which is a normal state period in which no signs of failure have yet appeared in the feature value. The selected feature value and information about the normal interval are stored in the fault prediction model condition storage section 124 as a condition for extracting learning data for machine learning.

[0100] Next, in step S107, the data extraction section 114 extracts data for machine learning from the feature value storage section 122 based on the extraction conditions stored in the fault prediction model condition storage section 124, and outputs the data for machine learning to the fault prediction model generation section 115. That is, the data of the selected feature value in the normal interval is extracted and output. This serves as a learning data extraction step.

[0101] Next, in step S108, the fault prediction model generating section 115 generates a learned model (i.e., a fault prediction model) by using the data for machine learning input from the data extracting section 114, and stores the learned model in the fault prediction model storing section 125. This serves as a learned model generating step.

[0102] The learned model (ie, the fault prediction model) can be generated by performing the above-described series of processes.

[0103] Next, a procedure of a process performed by the failure prediction device 100 by using the generated post-learning model (ie, failure prediction model) when determining whether the occurrence of a failure of the mechanical equipment 10 is imminent will be described.

[0104] Figure 8 1 is a flowchart illustrating the processing procedure. For example, the process for determining whether the occurrence of a failure of the mechanical equipment 10 is imminent is started by a user who instructs the start of the process by using the input section 140 of the failure prediction device 100. Alternatively, the control program of the failure prediction device 100 may be configured so that the process is automatically started according to the operation time of the mechanical equipment 10.

[0105] When the process starts, in step S201 , the sensor data collecting section 111 of the failure prediction apparatus 100 obtains measurement data from the sensor 11 that measures the state of the mechanical equipment 10 , and stores the measurement data in the sensor data storing section 121 .

[0106] Next, in step S202 , the feature value extraction section 112 extracts a feature value indicating a characteristic of the operating condition of the mechanical equipment 10 based on the sensor data stored in the sensor data storage section 121 , and stores the feature value in the feature value storage section 122 .

[0107] Next, in step S203, the data extraction section 114 obtains the information stored in the fault prediction model condition storage section 124 (i.e., information on the kind of feature values ​​used for machine learning), and extracts feature values ​​of the same kind as the feature values ​​used for machine learning from the feature value storage section 122. The extracted feature values ​​are output as evaluation data to the fault determination section 116. This serves as an evaluation feature value extraction step.

[0108] Next, in step S204, the fault determination part 116 inputs the feature value data for evaluation input from the data extraction part 114 into the learned model (ie, the fault prediction model) stored in the fault prediction model storage part 125, and calculates the degree of deviation between the input and the output.

[0109] Next, in step S205 , the failure determination section 116 compares the calculated degree of deviation with a determination threshold value, thereby determining whether the occurrence of a failure of the mechanical equipment 10 is imminent (ie, whether there is a sign of a failure).

[0110] In a case where the degree of deviation is equal to or greater than the determination threshold, that is, in a case where the result of step S205 is YES, it is determined that the occurrence of a failure in the mechanical equipment 10 is imminent, and the process proceeds to step S206 .

[0111] In step S206, the fault determination part 116 issues a notification instruction to the fault notification part 117. The fault notification part 117, which has received the notification instruction, notifies the user of the determination result of the fault determination part 116. When the notification is performed, in addition to performing the notification to the user through the user interface, the information related to the determination can also be stored in the storage part 120 or provided to the external device through the external interface. In order to perform the notification to the user, a process such as displaying the determination result on the display part 130 of the fault prediction device 100, outputting a voice message, or printing the determination result on a medium (such as paper) can be performed. When the notification to the user is completed, the process is completed.

[0112] In the case where the degree of deviation is less than the determination threshold, that is, in the case where the result of step S205 is no, it is determined that the occurrence of a failure of the mechanical equipment 10 is not imminent (that is, it is determined that there is no sign of a failure), and the processing ends. It should be noted that even in the case where it has been determined that there is no sign of a failure, the result can be notified to the user, and information related to the determination can be stored in a storage device or provided to an external device through an external interface.

[0113] As described above, according to the present embodiment, a feature value highly correlated with the occurrence of a fault can be selected from various feature values ​​indicating the state of the mechanical equipment based on the fault situation. Even in the case where only a small number of fault situations are collected, a fault model with high prediction accuracy can be generated. That is, even in the case of only a small number of fault situations, a feature value with high distinguishability for distinguishing a normal state from a fault state can be selected from various feature values ​​indicating the state of the mechanical equipment as data for machine learning. By performing machine learning by selecting feature values ​​highly correlated with the occurrence of a fault, a learned model that predicts the occurrence of a fault in advance with high accuracy can be generated.

[0114] Other embodiments

[0115] The embodiments of the present invention are not limited to the above-described embodiments and may be modified in various ways within the technical concept of the present invention.

[0116] For example, although a fault prediction model is generated by using an autoencoder by a so-called unsupervised learning method in the embodiment, the present invention can also be applied to a case where a fault prediction model is generated by using a so-called supervised learning method. Supervised learning is a method of constructing a model that predicts a result from an input (i.e., a learned model that inductively obtains a relationship between an input and an output) by providing a large data set of inputs and their results (in this case, labels) to a learning device and causing the learning device to learn the characteristics of these data sets.

[0117] In addition, although a method using a neural network has been described in the embodiments as an example of machine learning, the method of machine learning is not limited thereto, and for example, genetic programming, inductive logic programming, support vector machine, etc. may be used. Although a general-purpose calculator or a general-purpose processor may be used as a means for performing machine learning, high-speed processing may be performed by using a graphics processing unit having a GPGPU function, a large-scale PC cluster, etc.

[0118] In addition, although in the embodiment, the feature value selection section 131 selects feature values ​​to be used for machine learning based on the maximum value of the degree of separation for each feature value and the predetermined threshold value stored in the degree of separation threshold value setting section 132 in step S106, such selection may be performed without using a threshold value. For example, the feature values ​​may be arranged in order from a larger maximum value of the degree of separation to a smaller maximum value of the degree of separation, and feature values ​​up to a predetermined position (i.e., a predetermined number of feature values) may be selected for machine learning.

[0119] As described above, although according to the embodiment, a fault prediction model with high fault prediction accuracy can also be generated by using only data of a small number of fault conditions, it is not prohibited to perform additional machine learning when a new fault condition has occurred. That is, after the learned model is generated, additional learning can be performed by using data of the new fault condition. In this case, the feature value to be selected and the determination threshold can be changed with reference to the measurement data related to the fault condition to be subjected to additional learning.

[0120] The fault prediction device of the present invention can be applied to the fault prediction of various machines and equipment (such as industrial robots, service robots and processing machines) operated under the digital control of computers. The fault prediction system can be configured by integrating the mechanical equipment and the fault prediction device, or the fault generating device can be set as a part of the mechanical equipment.

[0121] The present invention can also be implemented by supplying a program that implements one or more functions of the embodiment to a system or device via a network or a recording medium and one or more processors of a computer of the system or device read out and execute the program. In addition, the present invention can also be implemented by a circuit (e.g., ASIC) that implements one or more functions.

[0122] Other embodiments

[0123] The embodiments of the present invention may also be implemented by a computer of a system or device and a method performed by the computer of the system or device, the computer reading and executing computer executable instructions (e.g., one or more programs) recorded on a storage medium (which may also be more completely referred to as a "non-transitory computer-readable storage medium") to perform the functions of one or more of the embodiments described above, and / or including one or more circuits (e.g., application specific integrated circuits (ASICs)) for performing the functions of one or more of the embodiments described above, the method by, for example, reading and executing computer executable instructions from the storage medium to perform the functions of one or more of the embodiments described above, and / or controlling the one or more circuits to perform the functions of one or more of the embodiments described above. The computer may include one or more processors (e.g., central processing unit (CPU), microprocessing unit (MPU)) and may include a network of independent computers or independent processors to read and execute the computer executable instructions. The computer executable instructions may be provided to the computer from, for example, a network or the storage medium. The storage medium may include, for example, a hard disk, a random access memory (RAM), a read-only memory (ROM), a storage device of a distributed computing system, an optical disk (such as a compact disk (CD), a digital versatile disk (DVD), or a Blu-ray disk (BD)). TM )), one or more of flash memory devices, memory cards, etc.

[0124] Other embodiments

[0125] The embodiments of the present invention may also be implemented by providing software (program) for performing the functions of the above-described embodiments to a system or device via a network or various storage media, and a computer or a central processing unit (CPU) or a microprocessing unit (MPU) of the system or device reads and executes the program.

[0126] While the present invention has been described with reference to exemplary embodiments, it is to be understood that the invention is not limited to the disclosed exemplary embodiments.The scope of the following claims is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures and functions.

Claims

1. A control method, comprising: obtaining a measurement value of a state of the mechanical equipment corresponding to a period of time during which the mechanical equipment reaches a second state from a first state; extracting at least two feature values ​​by using the measured values; For each of the at least two characteristic values, in a case where the predetermined time period is separated into a plurality of intervals at at least one timing, obtaining a separation degree, the separation degree indicating a degree of distinguishing an interval in which the mechanical device is in the first state from an interval in which the mechanical device is in the second state; selecting at least one eigenvalue from the at least two eigenvalues ​​based on a maximum value of the degree of separation of each of the at least two eigenvalues ​​in a predetermined period; setting a time period based on at least one timing at which the at least one characteristic value reaches a maximum value of the degree of separation; as well as Data for machine learning is extracted from data of the at least one feature value corresponding to the time period set in the setting.

2. The control method according to claim 1, wherein extracting data for machine learning comprises: setting a first time period based on the separation degree, the first time period being a time period during which the mechanical equipment is in the first state; as well as Data of the at least one feature value corresponding to the first time period is extracted as the data for machine learning.

3. The control method according to claim 1, wherein in obtaining the degree of separation, the data of the at least two characteristic values ​​along the timeline are separated into a first data set and a second data set at at least one timing in a predetermined time period, and the degree of separation is obtained based on the first mean value and the first variance of the first data set and the second mean value and the second variance of the second data set corresponding to each of the at least one timing in the predetermined time period.

4. The control method according to claim 3, further comprising: A screen is displayed to the operator, the screen being used to select whether the data of the at least one feature value corresponding to the time period set in the setting is used for the data for machine learning.

5. The control method according to any one of claims 1 to 4, further comprising: Using the data for machine learning to generate a learned model through machine learning; as well as The state of the mechanical equipment is determined by using the learned model.

6. The control method according to claim 5, wherein the determination of the state of the mechanical equipment comprises: inputting data of the at least one feature value corresponding to the time period set in the setting into the learned model; Obtaining a degree of deviation between input data input to the learned model and output data output from the learned model; setting a determination threshold based on a temporal change in the degree of deviation during a period in which the mechanical equipment reaches the second state from the first state; extracting a feature value of the same kind as the at least one feature value as an evaluation feature value by using a measurement value related to a state of the mechanical equipment corresponding to an evaluation time; as well as An index value indicating a degree of deviation of the mechanical equipment from the first state is obtained by using the evaluation feature value and the learned model, and a state of the mechanical equipment corresponding to the evaluation time is determined by using the index value and the determination threshold.

7. The control method according to claim 1, wherein the selecting the at least one characteristic value comprises: Each of the maximum values ​​of the degrees of separation is compared with a threshold value to select the at least one feature value.

8. The control method according to claim 1, wherein the selecting the at least one characteristic value comprises: A predetermined number of eigenvalues ​​having the largest maximum values ​​of the degree of separation are selected as the at least one eigenvalue.

9. The control method according to claim 1, The first period during which the mechanical equipment is in the first state is set to a period earlier than the earliest timing among the timings of the at least one characteristic value.

10. The control method according to claim 1, further comprising: Information related to the selection of the at least one characteristic value and / or information related to the extraction of data corresponding to a first period of time in which the mechanical equipment is in the first state is displayed on the display portion.

11. The control method according to claim 1, further comprising: Information related to the maximum value of the degree of separation, information related to the timing at which the degree of separation has reached the maximum value, and / or information related to a first period specifying that the mechanical equipment is in the first state is displayed on the display portion.

12. The control method according to claim 5, wherein generating a learned model comprises: The learned model is generated by machine learning using an autoencoder. 13 . The control method according to claim 5 , further comprising notifying, by a controller, a determination result of the state of the mechanical equipment.

14. A control device comprising a controller, wherein the controller is configured to: obtaining a measurement value of a state of the mechanical device corresponding to a period of time during which the mechanical device reaches a second state from a first state, by extracting at least two feature values ​​using the measured values, For each of the at least two characteristic values, in a case where the predetermined time period is separated into a plurality of intervals at at least one timing, obtaining a separation degree, the separation degree indicating a degree of distinguishing an interval in which the mechanical device is in the first state from an interval in which the mechanical device is in the second state; selecting at least one eigenvalue from the at least two eigenvalues ​​based on a maximum value of the degree of separation of each of the at least two eigenvalues ​​in a predetermined period; setting a time period based on at least one timing at which the at least one characteristic value reaches a maximum value of the degree of separation; as well as Data for machine learning is extracted from the data of the at least one feature value corresponding to the set time period.

15. Mechanical equipment comprising the control device according to claim 14.

16. A method of manufacturing a product by using a mechanical arrangement according to claim 15. 17 . A non-transitory computer-readable recording medium storing a control program for causing a computer to execute the control method according to claim 1 .

Images