Abnormality detection device, mechanical system, and abnormality detection method
Through the generation and learning of state and conditional feature quantities, abnormality and unknown degree are calculated, the error detection and omission problems when the operating conditions of the mechanical device are changed are solved, and the accuracy of abnormality detection is improved.
Patent Information
- Application Number
- CN202280094125.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-29
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, when the operating conditions of the mechanical device change, there are error judgment results such as error detection and omission, resulting in a decrease in the accuracy of abnormal detection.
The state signal generation unit, the conditional signal generation unit, the state feature quantity generation unit and the conditional feature quantity generation unit are used to calculate the abnormality and unknown degree through initial state learning and initial condition learning, and then perform abnormality determination to reduce error detection and omissions.
It improves the accuracy of abnormal detection of mechanical devices when operating conditions change, reduces the occurrence of error detection and omissions, and enhances the ability to determine abnormalities.
Smart Images

Figure CN119072667A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to detection of anomalies in mechanical devices. Background Art
[0002] Anomaly detection, which is to set sensors in mechanical devices, analyze signals from the sensors, and detect faults, degradation, etc. that occur in production equipment, is an important technology for achieving efficient operation of mechanical devices. By performing anomaly detection, when an abnormality occurs in a mechanical device due to long-term degradation, interference, etc. of the components of the mechanical device, the abnormality is detected and the operating conditions of the mechanical device can be changed, the mechanical device can be stopped, repaired, etc. Examples of components of the mechanical device include ball screws, reducers, bearings, pumps, etc. In addition, examples of abnormalities that occur in mechanical devices include increased friction, the occurrence of vibrations, damage to the frame, etc.
[0003] As an example of a technology for detecting anomalies, there are technologies called anomaly detection and outlier detection. In this anomaly detection technology, machine learning is performed to learn the characteristics of sensor signals in normal times and generate a model. Then, using the generated model, the degree of deviation of sensor signals obtained in a monitoring target time period that is the target of anomaly detection from sensor signals in normal times is quantitatively evaluated to detect anomalies.
[0004] According to this abnormality detection technology, there is an advantage that the occurrence of an abnormality can be detected even when the sensor signal at the time of the abnormality cannot be obtained in advance. On the other hand, when the operating conditions of the mechanical device when the sensor signal used for learning is obtained and the operating conditions of the time period of the monitored object are different, there is a problem of false detection due to the difference in operating conditions in this technology. Patent document 1 discloses the following technology, that is, the abnormality is calculated by machine learning, and the threshold value of the abnormality for judging normal or abnormal is adjusted using load data representing the load conditions of the mechanical device. The purpose of the technology described in patent document 1 is to improve the prediction accuracy of faults under changes in environmental conditions, load conditions, etc.
[0005] The control device described in Patent Document 1 obtains measurement values related to the state of the mechanical device and the load condition of the mechanical device when the mechanical device is in a normal state, uses the measurement values as learning data, and generates a trained model through machine learning. In addition, the control device described in Patent Document 1 obtains measurement values related to the state of the mechanical device from the normal state of the mechanical device to the abnormal state, and obtains the first threshold value using the obtained measurement values and the generated trained model.
[0006] Furthermore, the control device described in Patent Document 1 obtains the measured values related to the state of the mechanical equipment at the time of evaluation and the load condition of the mechanical equipment. Furthermore, the control device described in Patent Document 1 obtains the second threshold value based on the obtained load condition at the time of evaluation, the load condition at the time of training model generation, and the first threshold value. Furthermore, the control device described in Patent Document 1 determines the state of the mechanical equipment at the time of evaluation based on the trained model, the measured values related to the state of the mechanical equipment at the time of evaluation, and the second threshold value.
[0007] Furthermore, the control device of Patent Document 1 corrects the first threshold to the second threshold based on the difference in load conditions between when the learning model is generated and when it is evaluated, and reflects changes in the mechanical equipment between when the learning model is generated and when it is evaluated in the second threshold, thereby suppressing the occurrence of false detection.
[0008] In the control device described in Patent Document 1, the second threshold is not calculated accurately, and the determination result becomes inaccurate. In particular, the determination result becomes inaccurate when the change in state between the time of learning model generation and the time of determination does not appear due to the difference in load conditions.
[0009] As described above, the control device of Patent Document 1 has a problem that it is not possible to perform abnormality detection with less false detection when detecting the state of a mechanical device whose operating conditions change.
[0010] Patent Document 1: Japanese Patent Application Publication No. 2021-086220 Summary of the invention
[0011] As described above, when detecting the state of a mechanical device whose operating conditions change, there is a problem that abnormality detection with a small output of erroneous determination results such as false detection or omission cannot be performed.
[0012] The abnormality detection device involved in the present invention comprises: a state signal generating unit, which generates a state signal obtained by detecting the state of a mechanical device in a time series; a condition signal generating unit, which generates a condition signal obtained by detecting the operating condition representing the operating condition of the mechanical device in a time series; a state feature quantity generating unit, which generates a state feature quantity based on the state signal; a condition feature quantity generating unit, which generates a condition feature quantity based on the condition signal; and an initial state learning unit, which outputs the result obtained by learning based on the state feature quantity at the time of initial state learning, that is, the state feature quantity for initial learning as an initial state learning result. an initial condition learning unit, which outputs the result obtained by learning based on the conditional feature quantity at the time of initial condition learning, i.e., the conditional feature quantity for initial learning, as the initial condition learning result; an abnormality degree calculation unit, which obtains the initial state learning result or the additional state learning result as the state learning result, and calculates the abnormality based on the state learning result and the state feature quantity at the time of detection, i.e., the state feature quantity for detection; and an unknown degree calculation unit, which obtains the initial condition learning result or the additional condition learning result as the condition learning result, and calculates the unknown based on the condition learning result and the conditional feature quantity at the time of detection, i.e., the conditional feature quantity for detection.
[0013] The mechanical system involved in the present invention comprises: a mechanical device; a state signal generating unit, which generates a state signal obtained by detecting the state of the mechanical device in a time series; a condition signal generating unit, which generates a condition signal obtained by detecting the operating condition representing the operating condition of the mechanical device in a time series; a state feature quantity generating unit, which generates a state feature quantity based on the state signal; a condition feature quantity generating unit, which generates a condition feature quantity based on the condition signal; and an initial state learning unit, which uses the result obtained by learning based on the state feature quantity at the time of initial state learning, that is, the state feature quantity for initial learning as the initial state learning result. Output; an initial condition learning unit, which outputs the result obtained by learning based on the conditional feature quantity during initial condition learning, i.e., the conditional feature quantity for initial learning, as the initial condition learning result; an abnormality degree calculation unit, which obtains the initial state learning result or the additional state learning result as the state learning result, and calculates the abnormality based on the state learning result and the state feature quantity during detection, i.e., the state feature quantity for detection; and an unknown degree calculation unit, which obtains the initial condition learning result or the additional condition learning result as the condition learning result, and calculates the unknown based on the condition learning result and the conditional feature quantity during detection, i.e., the conditional feature quantity for detection.
[0014] The abnormality detection method involved in the present invention comprises: a state signal generating step, generating a state signal obtained by detecting the state of a mechanical device in a time series; a condition signal generating step, generating a condition signal obtained by detecting the operating condition representing the operating status of the mechanical device in a time series; a state feature quantity generating step, generating a state feature quantity based on the state signal; a condition feature quantity generating step, generating a condition feature quantity based on the condition signal; an initial state learning step, outputting the result obtained by learning based on the state feature quantity at the time of initial state learning, that is, the state feature quantity for initial learning, as an initial state learning result; an initial condition learning step, outputting the result obtained by learning based on the condition feature quantity at the time of initial condition learning, that is, the condition feature quantity for initial learning, as an initial condition learning result; an abnormality degree calculating step, obtaining the initial state learning result or the additional state learning result as the state learning result, and calculating the abnormality degree based on the state learning result and the state feature quantity at the time of detection, that is, the state feature quantity for detection; and an unknown degree calculating step, obtaining the initial condition learning result or the additional condition learning result as the condition learning result, and calculating the unknown degree based on the condition learning result and the condition feature quantity at the time of detection, that is, the condition feature quantity for detection.
[0015] Effects of the Invention
[0016] According to the present invention, when detecting the state of a mechanical device whose operating conditions change, abnormality detection with less output of erroneous determination results such as false detection and omission can be performed. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a diagram showing an example of the configuration of a mechanical system according to the first embodiment.
[0018] Figure 2 This is a diagram illustrating the configuration of a mechanical device and a control device according to the first embodiment.
[0019] Figure 3 This is a diagram showing a configuration example in which a processing circuit included in the mechanical system according to the first embodiment is constituted by a processor and a memory.
[0020] Figure 4 This is a diagram showing a configuration example in which a processing circuit included in the mechanical system according to the first embodiment is configured by dedicated hardware.
[0021] Figure 5 This is a diagram showing an example of the time waveforms of the motor speed and the motor torque according to the first embodiment.
[0022] Figure 6This is a diagram showing an example of the time waveforms of the motor speed and the motor torque in the continuous positioning according to the first embodiment.
[0023] Figure 7 This is a diagram showing an example of the time waveforms of the motor speed and the motor torque in the continuous positioning according to the first embodiment.
[0024] Figure 8 This is a diagram showing an example of an automatic encoder involved in implementation mode 1.
[0025] This is a diagram showing the time changes of the abnormality degree and the determination result of the conventional abnormality detection device involved in the first embodiment.
[0026] Fig. 9 This is a diagram showing an example of a configuration in which the abnormality determination unit is omitted from the abnormality detection device according to the first embodiment.
[0027] Fig.10 This is an example of the time change of the abnormality degree and the determination result generated by the configuration in which the abnormality determination unit is omitted from the abnormality detection device according to the first embodiment.
[0028] Fig.11 The abnormality degree, the unknown degree, and the time change of the determination result generated by the structure of the abnormality detection device according to the first embodiment without the abnormality determination unit are shown in FIG. Fig.10 Another example of being different.
[0029] Fig.12 This is a diagram showing the time changes of the abnormality degree, unknown degree, and judgment result generated by the abnormality detection device involved in the first embodiment.
[0030] Fig.13 This is a diagram showing an example of the operation flow of the abnormality determination unit according to the first embodiment.
[0031] Fig.14 This is a block diagram showing an example of the configuration of the mechanical system according to the first embodiment.
[0032] Fig.15 This is a block diagram showing an example of the configuration of a mechanical system according to the second embodiment.
[0033] Fig.16 This is a block diagram showing an example of the structure of the additional condition learning unit involved in the second embodiment.
[0034] Fig.17 This is a block diagram showing an example of the structure of the additional state learning unit involved in the second embodiment.
[0035] Fig.18This is a flowchart showing an example of the operation of the additional condition learning unit according to the second embodiment.
[0036] Fig.19 This is a flowchart showing an example of the operation of the additional state learning unit according to the second embodiment.
[0037] Fig. 20 This is a diagram showing an example of time changes in the abnormality degree, unknown degree, and judgment result generated by the configuration in which the additional condition learning unit and the additional state learning unit are omitted from the abnormality detection device involved in the second embodiment.
[0038] Fig.21 This is a diagram showing an example of time changes in the abnormality degree, unknown degree, and determination result generated by the configuration in which the additional state learning unit is omitted from the abnormality detection device according to the second embodiment. DETAILED DESCRIPTION
[0039] The embodiments are described in detail below based on the drawings. The embodiments described below are examples, and the scope of the present invention is not limited to the embodiments described below. The embodiments described below can be combined as appropriate.
[0040] Implementation Method 1
[0041] Figure 1 1 is a diagram showing an example of the structure of a mechanical system 100 according to the present embodiment. The mechanical system 100 includes: an abnormality detection device 1 that detects an abnormality occurring in a mechanical device 2; a mechanical device 2; and a control device 3 that controls the mechanical device 2. The abnormality detection device 1 includes a state signal generation unit 11 that generates a state signal ss, and a state feature quantity generation unit 12 that generates a state feature quantity sc. In addition, the abnormality detection device 1 includes: an initial state learning unit 13 that performs learning based on the state feature quantity sc and outputs an initial state learning result slr; and an abnormality degree calculation unit 14 that calculates the abnormality degree an.
[0042] In addition, the abnormality detection device 1 includes: a condition signal generating unit 15 that generates a condition signal cs; a condition feature quantity generating unit 16 that generates a condition feature quantity cc; and an initial condition learning unit 17 that performs learning based on the condition feature quantity cc and outputs an initial condition learning result clr. In addition, the abnormality detection device 1 includes: an unknown degree calculating unit 18 that calculates an unknown degree un; and an abnormality determining unit 19 that determines whether the state of the mechanical device 2 is abnormal or normal based on the abnormality degree an and the unknown degree un.
[0043] Here, the initial state learning performed by the initial condition learning unit 17 and the additional state learning described in Implementation Example 2 are each a mode of state learning. In addition, the detection state signal dss, the initial learning state signal lss, and the additional learning state signal alss described in Implementation Example 2 are each a mode of the state signal ss. The initial learning state feature quantity lsc and the detection state feature quantity dsc are each a mode of the state feature quantity. In addition, the initial state learning result slr and the additional state learning result aslr are each a mode of the state learning result. In addition, the initial state learning and the initial condition learning are sometimes referred to as initial learning.
[0044] The initial condition learning and the additional condition learning described in Implementation 2 are each a form of condition learning. The detection condition signal dcs, the initial learning condition signal lcs, and the additional learning state signal alss described in Implementation 2 are each a form of the condition signal cs. The initial learning condition feature quantity lcc and the detection condition feature quantity dcc are each a form of the condition feature quantity cc. The initial condition learning result clr and the additional condition learning result aclr described in Implementation 2 are each a form of the condition learning result.
[0045] Figure 1 The mechanical device 2 includes: a motor 20 that generates a driving force df; and a mechanical component 21 that is driven by the driving force df. The control device 3 includes: a command generation unit 30 that outputs an operating condition oc; and a control unit 31 that outputs the electric power pw toward the mechanical device 2 based on the operating condition oc. Examples of the mechanical device 2 include electronic component mounting machines, semiconductor manufacturing equipment, industrial robots, food manufacturing equipment, packaging machines, conveying equipment, automatic doors, press machines, roller feeders, air conditioners, generators, and the like.
[0046] The actions of the mechanical device 2 and the control device 3 are illustrated. The command generation unit 30 generates an operating condition oc, which becomes a control signal that specifies the action of the mechanical device 2. The command generation unit 30 supplies the electric power pw to the mechanical device 2 based on the operating condition oc. The motor 20 uses the electric power pw to generate a driving force df for the mechanical component 21 to drive the mechanical device 2. The mechanical component 21 can be any component that is moved by the driving force df. As examples of the mechanical component 21, a movable component that is moved by the driving force df of the motor 20, a component that connects movable components, etc. can be cited.
[0047] Figure 2 It is a diagram illustrating the configuration of the mechanical device 2 and the control device 3 according to the present embodiment. Figure 2 The mechanical device 2 shown has a ball screw 201, a coupling 202, a servo motor shaft 203 and a Figure 1 The servo motor 204 corresponds to the motor 20. Here, Figure 1 The driving force df is Figure 2 In the example of , it corresponds to the driving torque generated by the servo motor 204. The ball screw 201 has a movable part 2011 that moves when a ball screw shaft 2013 rotates, a guide part 2012 that restricts the direction of movement of the movable part 2011, and the ball screw shaft 2013.
[0048] The ball screw shaft 2013 and the servo motor shaft 203 are mechanically connected to the coupling 202. The driving torque generated by the servo motor 204, that is, the driving force df, is transmitted from the servo motor shaft 203 to the ball screw shaft 2013 via the coupling 202. The ball screw 201 converts the rotational motion into a linear motion through the screw mechanism, so that the movable part 2011 is Figure 2 The guide portion 2012 assists the movable portion 2011 in the state where it can move in the direction of the arrow so as to restrict the movement, thereby improving the accuracy of the movement of the movable portion 2011. In addition, the movable portion 2011 is connected to the mechanical component 21 (not shown), and the mechanical component 21 moves according to the purpose of the mechanical device 2.
[0049] Figure 2 The command generation unit 30 shown in FIG. 1 includes a PLC (Programmable Logic Controller) 301. The PLC 301 generates a command for operating the servo motor 204 and outputs it to a driver 311. Examples of commands include signals for indicating the position, speed, torque, etc. of the servo motor 204. Figure 1 The operating conditions οc correspond to the. If necessary, a PC (Personal Computer) 401 can be further set in the PLC 301. Figure 2 In the example of FIG. 4 , the PC 401 outputs a command related to the operation of the mechanical device 2 to the PLC 301. As the PC 401, for example, an industrial PC (Factory Automation PC or Industrial PC) can be used.
[0050] The control unit 31 includes a driver 311 and a current sensor 310. In addition, an encoder 205 for measuring the rotation angle of the servo motor 204 is installed in the mechanical device 2. The current sensor 310 measures the drive current supplied from the driver 311 to the servo motor 204. The drive current is Figure 1 Corresponding to the power pw.
[0051] The driver 311 performs feedback control on the servo motor 204 based on the measured value of the current sensor 310 and the measured value of the encoder 205, and supplies the drive current to the servo motor 204. In other words, the driver 311 performs feedback control so that the action of the servo motor 204 follows the command generated by the PLC 301. As described above, the command generated by the PLC 301 is Figure 1 Corresponding to the operating conditions oc.
[0052] In addition, Figure 1 , Figure 2 In the example of , the state signal ss used to detect the state of the mechanical device 2 is the motor torque mt. The state signal ss of the present embodiment is not limited to this example. The state signal ss only needs to be a signal containing information related to the state of the mechanical device 2. Examples of the state signal ss include signals obtained by measuring physical quantities such as position, speed, acceleration, current, voltage, torque, force, pressure, sound, and light intensity by sensors provided on the mechanical device 2 or sensors provided around the mechanical device 2. In addition, the state signal ss may be a signal of image information.
[0053] In addition, Figure 2 In the illustration, as examples of sensors used to obtain the state quantity sa, encoders 205 and current sensors 310 are illustrated, but the sensors used to obtain the state quantity sa are not limited to this. As examples of sensors used to obtain the state quantity sa, laser displacement meters, angle encoders, gyro sensors, vibrometers, acceleration sensors, voltmeters, torque sensors, pressure sensors, microphones, optical sensors, cameras, etc. can be cited. These sensors do not necessarily need to be arranged close to the mechanical device 2, the motor 20, etc., and can be arranged at any place if the state quantity sa can be generated. For example, it can be configured to arrange the acceleration sensor on the outer surface of the guide part 2012, and output the acceleration measured by the acceleration sensor as the state signal ss. Here, the outer surface of the guide part 2012 is the surface on the opposite side to the side where the movable part 2011 is arranged.
[0054] In addition, if Figure 2 As shown, the instruction generation unit 30 may have a PLC display 402 for displaying the status of the PLC 301, and may have a PC display 403 for displaying the status of the PC 401. In addition, a plurality of drive sources such as the servo motor 204 may be provided for one mechanical device 2. In addition, a plurality of drivers 311 may be provided for one mechanical device 2 as required. In addition, a single PLC 301 may centrally operate the mechanical device 2, or a plurality of PLCs 301 may coordinate to operate the mechanical device 2. The above is related to Figure 2The following is a description of the exemplary mechanical device 2 and the control device 3 shown.
[0055] In addition, the mechanical device 2 that becomes the target of abnormality detection by the abnormality detection device 1 is not limited to the mechanical device 2 in the Figure 2 In addition, the types of abnormalities that become the object of abnormality detection are not limited to those in Figure 2 The abnormality detection device 1 can be widely used for all phenomena occurring in the mechanical device 2. Events occurring in the mechanical device 2, phenomena occurring in the mechanical device 2, the state of the mechanical device 2, etc. can be regarded as abnormal. Examples of phenomena that can be regarded as abnormal include the mixing of foreign matter into the mechanical device 2, damage to the frame of the mechanical device 2, deterioration of grease, peeling of materials, defects in workpieces, problems with fluids, poor installation of the device, poor assembly, etc. In addition, a situation in which two or more phenomena occur can also be detected as abnormal.
[0056] Figure 3 11 is a diagram showing a configuration example in which the processing circuit included in the mechanical system 100 according to the present embodiment is composed of a processor 1151 and a memory 1152. For example, Figure 1 The abnormality detection device 1, the control device 3, Figure 2 The driver 311 etc. shown may have Figure 3 processing circuit. In the case where the processing circuit is composed of a processor 1151 and a memory 1152, each function of the processing circuit of the abnormality detection device 1, the control device 3, the driver 311, etc. is implemented by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and stored in the memory 1152. In the processing circuit, the program stored in the memory 1152 is read out and executed by the processor 1151, thereby realizing each function. That is, in the case where the abnormality detection device 1, the control device 3, the driver 311, etc. have a processing circuit, the processing circuit has a memory 1152, and the memory 1152 is used to store the program that the processing of the abnormality detection device 1, the control device 3, the driver 311, etc. is finally executed. In addition, these programs can be said to cause the computer to execute the order and method executed by the abnormality detection device 1, the control device 3, the driver 311, etc.
[0057] Here, the processor 1151 may be a computing unit called a CPU (Central Processing Unit), a processing device, a computing device, a microprocessor, a microcomputer, or a DSP (Digital Signal Processor). The memory 1152 may be, for example, a nonvolatile or volatile semiconductor memory such as a RAM, a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable ROM), or an EEPROM (registered trademark) (Electrically EPROM). In addition, the memory 1152 may be configured as a storage unit such as a magnetic disk, a floppy disk, an optical disk, a compact disk, a mini disk, or a DVD (Digital Versatile Disc).
[0058] Figure 4 1 is a diagram showing a configuration example in which the processing circuit included in the mechanical system 100 according to the present embodiment is configured by dedicated hardware. Figure 1 The abnormality detection device 1, the control device 3, Figure 2 The driver 311 shown, etc., may have Figure 4 When the processing circuit is composed of dedicated hardware, Figure 4 The processing circuit 1161 shown can be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. The processing circuit of the mechanical system 100 can be implemented by the processing circuit 1161 for each function. Figure 1 The multiple functions of the abnormality detection device 1, the control device 3, the driver 311, etc. shown in the figure can also be realized by integrating the multiple functions through the processing circuit 1161. In addition, the abnormality detection device 1, the control device 3, the driver 311, the PLC 401, etc. can also be connected via a network. In addition, at least any one of the abnormality detection device 1, the control device 3, the driver 311, the PLC 401, etc. can exist on the cloud server.
[0059] In addition, it is also possible to Figure 2The structure omits the driver 311, PLC 301, PC 401, etc. For example, instead of the driver 311, PLC 301, PC 401, etc., a device for implementing the abnormality detection of the abnormality detection device 1 may be prepared separately, and the operation of the abnormality detection device 1 may be performed by the device. For example, a device having a battery, a microcomputer, a sensor, a display, and a communication function may be prepared, and the device may obtain the sound emitted by the mechanical device 2 through a microphone as a state signal ss. Furthermore, the abnormality detection device 1 may detect the state of the mechanical device 2 based on the state signal ss.
[0060] In addition, Figure 2 In the configuration example, the PLC display 402, the PC display 403, etc. can also be omitted. Figure 2 In the structural example, these display devices can be replaced by using LEDs or the like of the driver 311 and PLC 301 to display the states of the servo motor 204, the ball screw 201, etc. In addition, the normal and abnormal determination results related to the state can be displayed using LEDs or the like. In addition, it is also possible to configure such that the state and the determination results related to the state are not displayed, and when it is determined that an abnormality has occurred, the driving of the servo motor 204 is stopped.
[0061] In addition, Figure 2 In the description of FIG. 2 , the speed of the servo motor 204 is obtained from the encoder 205 provided in the servo motor 204, but the present embodiment is not limited to the above-described method. Figure 1 In the structure, the control signal from the instruction generation unit 30 that issues a drive command to the motor 20 can be used as the operation condition oc, and the condition signal generation unit 15 generates the motor speed as the condition signal cs.
[0062] In addition, Figure 2 In the example of the servo motor 204, the servo motor 204 is a rotary type servo motor. Figure 1 The abnormality detection device 1 of this embodiment can also be applied when the motor 20 is set as a motor other than a rotary type. Examples of motors other than rotary types include linear servo motors, induction motors, stepping motors, brushed motors, ultrasonic motors, etc. In addition, the abnormality detection device 1 of this embodiment can also be applied when a drive source different from the motor is used instead of the motor. As an example of a drive source other than a motor, for example, the mechanical device 2 can also be driven by an internal combustion engine such as a gasoline engine, a jet engine, a rocket engine, and a gas turbine. As described above, the drive source is not limited to being driven by electricity.
[0063] Furthermore, the mechanical device 2 may also be driven by natural energy such as wind power, geothermal energy, and water power. For example, the mechanical device 2 may be a wind power generation device, a geothermal power generation device, a water power generation device, and the like. When the mechanical device 2 is driven according to an instruction to an electric motor, an internal combustion engine, and the like, the instruction can be used as an operating condition OC. The instruction does not contain a large amount of interference, so the instruction is used as an operating condition OC, thereby enabling the abnormality detection device 1 to detect an abnormality with high accuracy. In addition, the abnormality detection device 1 can suppress the occurrence of false detection, omission, and the like and detect an abnormality. When the mechanical device 2 is driven by natural energy, such as a wind power generation device, the mechanical system 100 sometimes does not include the control device 3.
[0064] In addition, the example Figure 1 In the mechanical device 2, the ball screw 201 and the coupling 202 are included as structural elements, but the structural elements of the mechanical device 2 are not limited to these elements. Examples of structural elements of the mechanical device 2 other than the ball screw 201 and the coupling 202 include a speed reducer, a guide, a transmission belt, a screw, a pump, a bearing, a frame, etc. As described above, the abnormality detection device 1 can be applied to various mechanical devices 2.
[0065] The operation of the abnormality detection device 1 is illustrated. Figure 2 In the example, the increase in vibration and friction caused by the deterioration of the sliding part of the ball screw 201 are taken as examples of abnormalities detected by the abnormality detection device 1. The state signal generating unit 11 obtains the state quantity sa obtained by detecting the physical phenomenon occurring in the mechanical device 2 using sensors, etc., and outputs it as a time series state signal ss. Here, the physical phenomenon is a quantity that can be detected using sensors, etc. on the mechanical device 2. For example, the state quantity sa can be set to a quantity that is manifested by the influence of a failure, degradation, etc. occurring in the mechanical device 2. The state quantity sa can also be set to a quantity that can detect the state of the mechanical device 2, the failure, degradation, etc. occurring in the mechanical device 2. In addition, the structure of the abnormality detection device 1 can also be configured as a quantity that is correlated with the state quantity sa and the state of the mechanical device 2, the failure, degradation, etc. occurring in the mechanical device 2.
[0066] In this embodiment, the time series signal is set as a signal having information associated with each of the multiple time points. For example, if a time point among the multiple time points of the time series signal is specified, the signal corresponding to the time point or the signal determined by the value shown by the signal can be used as the time series signal.
[0067] Here, the state signal ss obtained by the state signal generating unit 11 detecting the state of the mechanical device 2 during the initial state learning time is referred to as the initial learning state signal lss. Here, the initial state learning time is preferably the time when the mechanical device 2 is in a normal state. In addition, the time of the object of abnormal detection during which the state signal generating unit 11 detects the state of the mechanical device 2 is set as the detection time. Moreover, the state signal ss detected by the state signal generating unit 11 during the detection time is referred to as the detection state signal dss. In addition, the relationship between the initial state learning time and the detection time is not limited, but when the detection time is later than the initial state learning time, there is an advantage that the result of the initial state learning can be used for abnormality detection.
[0068] exist Figure 2 In the example of , the value of the driving torque calculated based on the value of the current flowing in the servo motor 204 as an example of the motor 20 is set as the state quantity sa. In addition, the state signal generating unit 11 measures the value of the driving current using the current sensor 310, converts the current value into the value of the driving torque generated by the servo motor 204, that is, the torque value, and sets it as the state signal ss. The state quantity sa can change every moment according to the behavior of the mechanical device 2, time, etc.
[0069] In addition, the state feature quantity generating unit 12 obtains the state signal ss in a time series. And, the state feature quantity sc is generated based on the state signal ss of the time series. The state feature quantity sc is preferably a quantity obtained by extracting the characteristics representing the state of the mechanical device 2. The state feature quantity sc may not be a signal of a time series, but is preferably generated in a time series. In addition, the state feature quantity generating unit 12 may generate the state feature quantity sc one by one for each group including a plurality of time points at which the state signal ss is generated. In addition, the state feature quantity generating unit 12 may generate the state feature quantity sc one by one for each plurality of time points at which the state signal ss is generated. Here, the state feature quantity sc generated by the state feature quantity generating unit 12 based on the initial learning state signal lss is referred to as the initial learning state feature quantity lsc. In addition, the state feature quantity sc generated by the state feature quantity generating unit 12 based on the detection state signal dss is referred to as the detection state feature quantity dsc.
[0070] The learning result of the initial state learning unit 13 performing learning based on the initial learning state feature quantity lsc is output as the initial state learning result slr. The learning performed by the initial state learning unit 13 is called initial state learning. For example, the initial state learning unit 13 can generate a model about the characteristics of the initial learning state signal lss, and output the structure, parameters, etc. of the model as the initial state learning result slr. In addition, in the abnormality detection device 1, instead of the initial state learning unit 13, a learning model after performing initial state learning, an output of the completed initial state learning result slr, etc. can be provided. For example, a model based on the initial state learning result slr output by the initial state learning unit 13 described in the present embodiment can be provided. In the case of having a trained learning model, an output of the completed initial state learning result slr, etc., it is possible to use the result of the initial state learning instead of performing initial state learning.
[0071] The abnormality calculation unit 14 calculates the abnormality an based on the detection state feature quantity dsc and the initial state learning result slr. The abnormality calculation unit 14 can calculate the degree of deviation between the characteristics of the detection state signal dss and the characteristics of the initial learning state signal lss as the abnormality an. In addition, the abnormality calculation unit 14 can calculate the difference between the characteristics of the detection state feature quantity dsc and the characteristics of the initial learning state feature quantity lsc as the abnormality an.
[0072] In addition, for example, the initial state learning result slr includes the structure of the model, the parameters of the model, etc. In this case, the abnormality calculation unit 14 generates a model based on the structure of the model, the parameters of the model, etc. In addition, the abnormality calculation unit 14 can calculate the difference between the output when the initial learning state feature lsc is input to the model and the output when the detection state feature dsc is input to the model as the abnormality an.
[0073] The condition signal generating unit 15 obtains the operating condition of the mechanical device 2 as the operating condition oc and generates the condition signal cs. The operating condition oc only needs to indicate the operating status of the mechanical device 2. Figure 1 , Figure 2 In the description of FIG. 1 , the operating condition oc is set to a set value or command value of the speed of the motor 20, a set value or command value of the acceleration of the motor 20, a set value or command value of the moving distance of the motor 20, etc. Figure 1 , 2 In the description of , the condition signal cs is set as a speed command, and the command speed ds is a time series signal. In addition, the condition signal cs is set as a signal for obtaining the operating condition oc in a time series. Other examples of the operating condition oc and the condition signal cs include a jump, a load, an external air temperature, a pressure, a flow rate, and the like.
[0074] Here, the condition signal cs obtained by the condition signal generating unit 15 detecting the state of the mechanical device 2 during the initial condition learning time is referred to as the initial learning condition signal lcs. Here, the initial condition learning time is preferably the time when the mechanical device 2 is in a normal state. In addition, the condition signal cs detected by the condition signal generating unit 15 during the detection time, which is the time of the detection target of the state of the mechanical device 2, is referred to as the detection condition signal dcs.
[0075] The relationship between the initial condition learning time and the detection time is not limited, but the result of the initial condition learning can be used for abnormality detection, so the detection time is preferably a time later than the initial condition learning time. Here, the initial state learning time and the initial condition learning time do not necessarily have to be consistent. On the other hand, the detection time in the description of the detection state signal dss is consistent with the detection time in the description of the detection condition signal dcs.
[0076] The condition feature quantity generating unit 16 generates the initial learning condition feature quantity lcc based on the initial learning condition signal lcs. For example, the condition feature quantity generating unit 16 can extract the feature representing the characteristics of the operating condition oc during the initial condition learning time from the initial learning condition signal lcs, and generate the initial learning condition feature quantity lcc. The initial condition learning unit 17 performs learning based on the initial learning condition feature quantity lcc, and outputs the learning result as the initial condition learning result clr. The learning performed by the initial condition learning unit 17 is called initial condition learning. Initial condition learning, initial state learning, etc. are sometimes referred to as initial learning.
[0077] For example, the initial condition learning unit 17 can model the characteristics of the operating condition OC during the initial condition learning time based on the condition feature quantity lcc for initial learning, and output the structure and parameters of the model as the initial condition learning result clr. In addition, in the abnormality detection device 1, a trained learning model, an output initial condition learning result clr, etc. can be replaced with the initial condition learning unit 17. In the case where the abnormality detection device 1 has a trained learning model, an output initial condition learning result clr, etc., it is possible to perform high-precision abnormality detection in a short time without performing learning, but using the results of learning. In addition, the calculation load can be reduced. For example, the trained learning model can be a model based on the initial condition learning result clr output by the initial condition learning unit 17.
[0078] The unknown degree calculation unit 18 calculates the unknown degree un based on the initial condition learning result clr and the detection condition signal dcs. The unknown degree un can be a quantity indicating the degree of deviation between the initial learning condition signal lcs and the detection condition signal dcs. In addition, the unknown degree calculation unit 18 can calculate the degree of deviation between the characteristics of the detection condition feature quantity dcc and the initial learning condition feature quantity lcc as the unknown degree un.
[0079] In addition, for example, the initial condition learning result clr includes the structure of the model, the parameters of the model, etc. In this case, the unknown degree calculation unit 18 generates a model based on the structure of the model, the parameters of the model, etc. Furthermore, the unknown degree calculation unit 18 can calculate the difference between the output when the initial learning condition feature quantity lcc is input to the model and the output when the detection condition feature quantity dcc is input to the model as the unknown degree un.
[0080] The abnormality determination unit 19 determines whether an abnormality has occurred in the mechanical device 2 based on the abnormality degree an and the unknown degree un, and outputs the result as a determination result jr. For example, the abnormality determination unit 19 may determine that the state of the mechanical device 2 is abnormal when the abnormality degree an is greater than a predetermined first threshold value and the unknown degree un is less than a predetermined second threshold value.
[0081] Furthermore, in any of the two cases, when the abnormality an is less than or equal to the first threshold value, and when the unknown degree un is greater than or equal to the second threshold value, the abnormality determination unit 19 can determine that the state of the mechanical device 2 is normal. Here, the case where the abnormality an is less than or equal to the first threshold value is the case where the abnormality an is less than or equal to the first threshold value. In addition, the case where the unknown degree un is greater than or equal to the second threshold value is the case where the unknown degree un is greater than or equal to the second threshold value.
[0082] In addition, the abnormality detection device 1 can also be set to a mode without the abnormality determination unit 19. In this case, a device outside the abnormality detection device 1 can execute the processing of the abnormality determination unit 19. In addition, the processing of the abnormality determination unit 19 can be executed by the operator. In addition, the abnormality determination unit 19 performs machine learning on the abnormality degree an and the unknown degree un and generates a model as learning data. Moreover, the abnormality determination unit 19 can make a judgment based on the generated model and the abnormality degree an and the unknown degree un obtained during the detection time of the object of abnormality detection of the mechanical device 2. In addition, the abnormality detection device 1 can have a display unit that displays the judgment result jr, the unknown degree un, the abnormality degree an, etc.
[0083] Figure 51 is a diagram showing an example of the time waveform of the motor speed ms and the motor torque mt according to the present embodiment. The condition signal generating unit 15 generates the command speed ds as the condition signal cs. Figure 5 In (a), the time series waveform of the command speed ds generated by the condition signal generating unit 15 as the condition signal cs is represented by a dotted line or a dashed line. The command speed ds is the command speed ds of the motor specified by the operating condition oc generated by the command generating unit 30. Figure 5 In (a), the motor speed ms calculated from the measurement result of the encoder 205 based on the command speed ds is represented by a solid line. Figure 5 (b) shows a time-series waveform of the motor torque mt generated by the state signal generating unit 11 as the state signal ss. Figure 5 (a) and Figure 5 The horizontal axis of (b) represents time. Figure 5 (a) A point with a label on the time axis and a point with the same label as the point in the time axis Figure 5 The points on the time axis of (b) specify the same moment in time.
[0084] right Figure 5 The time variation of the command speed ds shown in (a) is described. Between the time tr0 and the time tr1, the command speed ds is 0. Between the time tr1 and the time tr2, the command speed ds becomes a command to accelerate at a constant acceleration. Moreover, the command speed ds, which is 0 at the time tr1, reaches the speed Vcmd at the time tr2. Between the time tr2 and the time tr3, the command speed ds remains at the speed Vcmd. Between the time tr3 and the time tr4, the command speed ds becomes a command to decelerate at a constant acceleration. The command speed ds, which is the speed Vcmd at the time tr3, becomes 0 at the time tr4. Moreover, between the time tr4 and the time tr5, the command speed ds remains at 0.
[0085] Figure 5 The motor speed ms in (a) is the actual measured value of the speed of the motor 20. In other words, the motor speed ms is Figure 2 The servo motor 204 is controlled by the command generation unit 30, that is, by Figure 2 The driver 311 controls the motor speed ms so that the motor speed ms follows the command speed ds. The relationship between the command speed ds and the motor speed ms varies depending on the structure of the command generating unit 30. Figure 5 In (a), the motor speed ms follows the command speed ds with a slight delay. Figure 5In (b), the time series waveform of the motor torque mt calculated from the measured value of the current sensor 310 is represented by a solid line. The signal directly obtained from the current sensor 310 is assumed to be Figure 2 The state signal generating unit 11 converts the three-phase current to generate a signal (three-phase current, not shown) by measuring the three-phase current flowing in the servo motor 204. Figure 5 (b) The motor torque mt shown is used as the state signal ss.
[0086] like Figure 5 As shown in the example of FIG. 1 , a sensor can be appropriately provided for the mechanical device 2. In addition, the sensor can be a structure included in the state signal generating unit 11 or a structure included in the abnormality detection device 1. In addition, the state signal generating unit 11 can appropriately perform a conversion on the measurement value of the sensor. Figure 5 The state signal generating unit 11 generates the state signal ss based on the measurement result of the sensor that detects the value of the torque, force, vibration, speed, position, light quantity, sound, etc. generated in the mechanical component 21, or instead of the sensor. In addition, as examples of sensors, torque sensors, force sensors, vibration sensors, gyro sensors, encoders, laser displacement meters, optical sensors, microphones, etc. can be cited. In addition, the state signal generating unit 11 can directly generate the current value of the three phases detected by the current sensor 310 as the state signal ss. In addition, the state signal generating unit 11 can generate the rotation angle of the servo motor 204 obtained from the encoder 205, or the motor speed ms obtained by numerical differentiation or the like from the rotation angle as the state signal ss.
[0087] The time series changes of the state signal ss are described below. Figure 5 The time series waveform of the motor torque mt described in (a) is a waveform indicating the acceleration of the servo motor 204 from time tr1 to time tr2. And, from time tr3 to time tr4, it is a waveform indicating the deceleration of the servo motor 204. Correspondingly, Figure 5 The motor torque mt shown in (b) increases from time tr1 to time tr2. Figure 5 The motor torque mt shown in (b) decreases between the time tr3 and the time tr4.
[0088] exist Figure 5The time waveform (b) shows that the friction force acts and the motor torque mt changes depending on the motor speed ms. For example, from time tr1 to time tr2, the motor speed ms is accelerated at the same acceleration, but as the motor speed ms increases, the motor torque mt increases.
[0089] exist Figure 5 (a) and Figure 5 (b) shows an example of a waveform when a single drive called positioning is performed from a stopped state of the servo motor 204. Figure 5 In the example of , the number of positioning is shown as one, but multiple positioning can also be performed. Figure 5 In the embodiment, the operation of positioning the mechanical part 21 by the servo motor 204 is described, but the application of the abnormality detection device 1 of this embodiment is not limited to the positioning operation. For example, the abnormality detection device 1 of this embodiment can also be applied to the control different from the positioning operation, such as speed control and torque control. In addition, for example, the abnormality detection device 1 of this embodiment can also be applied to the abnormality detection of the mechanical device 2 that is not controlled according to the instruction.
[0090] In addition, Figure 5 In the example of , the case where the time when the command speed ds becomes constant is included is illustrated. Figure 5 In the example of , the case where the time waveform of the command speed ds is in a trapezoidal shape is illustrated. The abnormality detection device 1 of this embodiment can also be applied to the case where the time waveform of the command speed ds is not in a trapezoidal shape when there is no time when the command speed ds becomes constant, that is, when the waveform becomes a triangle. In addition, Figure 5 In the example of , the slope of the speed during acceleration is constant, that is, the acceleration of the command speed ds becomes a shape close to a rectangle. However, the structure that can be applied to the abnormality detection device 1 is not limited by the waveform of the time series of the instruction. For example, in order to suppress the occurrence of vibrations associated with rapid acceleration, the abnormality detection device 1 of this embodiment can also be applied to the waveform of the acceleration of the command speed ds when the control of imposing an upper limit on the jump is executed. In addition, in order to suppress the vibration of the mechanical device 2, the abnormality detection device 1 of this embodiment can also be applied to the structure that implements filtering processing on the instruction.
[0091] In addition, Figure 2In the example of FIG, the instruction to the servo motor 204 generated by the PLC 301 is used as the operating condition OC, but the operating condition OC is not limited to the instruction to the motor 20. The operating condition OC only needs to include information related to the operation of the mechanical device 2. In addition, it is preferable to select an amount that has little correlation with the occurrence of an abnormality and affects the state signal ss or the state amount sa as interference as the operating condition OC.
[0092] The information included in the condition signal cs is preferably information that is unlikely to cause an abnormality in the mechanical device 2. Furthermore, the information included in the condition signal cs is preferably information that may affect the state signal ss as interference. The condition signal generating unit 15 is preferably configured to generate the condition signal cs.
[0093] In addition, it is preferred that the detected value of the operating condition oc or the condition signal cs does not change significantly between the state where the abnormality occurs and the state where the abnormality does not occur. In addition, even if an error occurs between the unknown degree un in the case of the abnormality and the unknown degree un in the case of the non-abnormality, it is preferred that the difference does not cause a change exceeding the threshold value set in the unknown degree un.
[0094] The abnormality determination unit 19 outputs the determination result of whether the mechanical device 2 is in an unknown state to the control device 3. When the determination result that the mechanical device 2 is in an unknown state is obtained, the command generation unit 30 outputs the operating condition oc for changing the mechanical device 2 to a known state. Moreover, after the abnormality determination unit 19 obtains the determination result that the mechanical device 2 is in a known state based on the value of the unknown degree un, it can perform abnormality detection based on the unknown degree un and the abnormality degree an. Here, the abnormality determination unit 19 can output the unknown degree un to the control device 3 instead of the determination result, and the control device 3 determines whether the mechanical device 2 is in an unknown state or a known state.
[0095] In the abnormality detection device 1, a display unit is provided for displaying a determination result jr of whether the mechanical device 2 is in an unknown state. When the mechanical device 2 is in an unknown state, the operator can perform an operation of switching to an operation condition oc that changes the mechanical device 2 to a known state. Here, instead of the determination result jr, the unknown degree un can be displayed on the display unit, and the operator can determine whether the mechanical device 2 is in an unknown state.
[0096] As examples of information included in the conditional signal CS, the outside air temperature, the vibration value associated with the mechanical device 2, the mass of the workpiece processed by the mechanical device 2, the input to the mechanical device 2 when the mechanical device 2 is operated, etc. can be cited. Here, the vibration value associated with the mechanical device 2 is a value related to the vibration that occurs when contacting at least a part of the mechanical device 2. In addition, the vibration value associated with the mechanical device 2 can be set to a value related to the vibration involved in causing at least a part of the mechanical device 2 to vibrate or excite vibration. As the above examples, the ground, the stand, the surrounding air, etc. on which the mechanical device 2 is installed can be cited. As examples of numerical values related to vibration, the amplitude, frequency and their combination of vibration can be cited. The conditional signal generating unit 15 can generate the conditional signal CS based on this information. Of course, multiple types of information can be combined to generate the conditional signal CS.
[0097] The operations of the state feature quantity generating unit 12 and the condition feature quantity generating unit 16 will be described. Figure 6 , Figure 7 FIG. 1 is a diagram showing an example of the time waveform of the command speed ds and the motor torque mt of the continuous positioning according to the present embodiment. Figure 6 (a) Figure 7 In (a), the time waveform of the command speed ds outputted by the condition signal generating unit 15 as the condition signal cs is shown. Figure 6 (a) shows the time waveform from time 0 seconds to time 10 seconds. Figure 7 (a) shows the time waveform from time 10 seconds to time 20 seconds. Figure 6 (b) Figure 7 (b) shows the time waveform of the motor torque mt outputted by the state signal generating unit 11 as the state signal ss. Figure 6 (b) shows the time waveform from time 0 seconds to time 10 seconds. Figure 7 (b) shows the time waveform in the range from time 10 seconds to time 20 seconds.
[0098] Figure 6 (a) Figure 6 (b) Figure 7 (a) and Figure 7 The horizontal axis of (b) represents time in seconds (s). Figure 6 (a) A position with a number on the time axis and a position with the same number as the number Figure 6 The positions on the time axis of (b) become the same time. Figure 7 (a) A position with a number on the time axis and a position with the same number as the number Figure 7 The positions on the time axis of (b) become the same time. Figure 6 (a) and Figure 7 The vertical axis of (a) represents the command speed ds, and the unit is rotation speed per minute (r / min (round per minute)). Figure 6 (b) and Figure 7 The vertical axis of (b) represents the motor torque mt, and the unit is Newton meter (Nm).
[0099] Figure 5 The figure shows one positioning, in contrast, Figure 6 and Figure 7 , 10 consecutive positionings are shown. The 10 positionings are referred to as positioning D1 to positioning D10. From positioning D1 to positioning D10, at least one of the commanded speed, acceleration during acceleration, acceleration during deceleration, and moving distance is different from each other. For example, at positioning D1, the maximum speed is 3200r / min, and at positioning D2, the maximum speed is 2200r / min.
[0100] In addition, at positioning D2, the acceleration during acceleration is large, and the acceleration during deceleration is smaller than that at positioning D1. In addition, at positioning D2, the absolute value of the acceleration during acceleration is smaller than that at positioning D1, and the absolute value of the acceleration during deceleration is larger than that at positioning D1. In addition, the direction of movement of the servo motor 204 at positioning D3 is different from that at positioning D1, and the direction of the speed becomes negative. In addition, at positioning D3, the moving distance is smaller than that at positioning D1 and positioning D2, and the shape of the time waveform of the command speed ds, i.e., the operating condition oc, becomes a triangle. As described above, the shapes of the command speeds ds from positioning D1 to positioning D10 are different from each other. Moreover, the operating condition oc generated by the command generation unit 30 is different from each other between positioning D1 and positioning D10.
[0101] The command generation unit 30 generates operating conditions OC in accordance with the state of the operation of the mechanical device 2. For example, when the mechanical device 2 is a conveying device, it is preferred to improve the efficiency of the conveying process, so the command generation unit 30 preferably generates operating conditions OC to complete each positioning in the shortest possible time. In addition, in the case where it is necessary to reduce shaking, impact, etc. when conveying a workpiece, the command generation unit 30 sets an upper limit on the speed, acceleration, jerk, etc., and generates operating conditions OC in a manner that the speed, acceleration, jerk, etc. do not exceed the upper limit. In addition, for example, when the mechanical device 2 is an electronic component mounting machine, in the case where the installation position of the electronic component is frequently changed, the command generation unit 30 generates operating conditions OC in which the moving distance and moving direction are different for each positioning.
[0102] The state characteristic quantity generating unit 12 generates the state characteristic quantity sc based on the motor torque mt. Figure 6 (b) and Figure 7(b) is shown. The state feature quantity generating unit 12 preferably generates the state feature quantity sc in such a manner that the number of variables of the state feature quantity sc becomes the same between the multiple state feature quantities sc generated in a time series. Here, the number of variables of the state feature quantity sc may be, for example, the number of variable parameters or the number of parameters possessed by each of the state feature quantities sc. The multiple parameters can take different values from each other between the multiple state feature quantities sc. By comparing the multiple state feature quantities sc, the change in the state of the mechanical device 2 is detected, so that the situation where the number of variables between the state feature quantities sc is the same can be more easily compared with high precision. The generated state feature quantity sc is input to the initial state learning unit 13 and the abnormality degree calculation unit 14, and learning is performed by the initial state learning unit 13 based on the state feature quantity sc. In addition, the calculation of the abnormality degree an is performed by the abnormality degree calculation unit 14.
[0103] For example, the state feature quantity generating unit 12 obtains the time series signal of the motor torque mt from the time ts1 to the time te1 at the time points of N equal time intervals as a set of state feature quantities sc1. The number of samples of the motor torque mt at this time, that is, the number of variables of the state feature quantity sc, is N. The period from the time ts1 to the time te1 is called the processing time. Specific numerical values are given as examples. For example, when the sampling period is 1 millisecond (ms), the time te1 is 1501ms, and the time ts1 is 1700ms, the state feature quantity sc becomes a vector with the number of variables N=100. The state feature quantity sc1 cited here is only an example, and the present embodiment is not limited to the above-mentioned method. The sampling period, the start time ts1 of the processing, the end time te1 of the processing, etc. can be appropriately changed. For example, N can be set to a large value to generate a set of state feature quantities sc across multiple positioning. In addition, time series signals such as the state quantity sa, the state feature quantity sc, the operating condition oc, and the condition feature quantity cc are not limited to signals with equal time intervals. For example, the time interval of the time series signal may be set short only in a portion where a large amount of data needs to be acquired.
[0104] The state feature quantity generating unit 12 sequentially generates the state feature quantity sc while changing the time set as the target. Figure 6 The motor torque mt from the time ts2 to the time te2 is set as the second group of state characteristic quantities sc2. At this time, in order to make the number of variables of the state characteristic quantity sc consistent with N, the value (te2-ts2) obtained by subtracting the time ts2 from the time te2, which is the interval between the start time and the end time of the processing, and the value (te1-ts1) obtained by subtracting the time ts1 from the time te1 can be the same value.
[0105] Similarly, the state feature quantity generating unit 12 generates the state feature quantity according to Figure 6 The state characteristic quantity sc3 is generated based on the motor torque mt obtained in the implementation of the positioning D3 shown in FIG. In addition, the state characteristic quantity sc4 is generated based on the motor torque mt obtained in the implementation of the positioning D4. In addition, based on the Figure 7 The state characteristic amount sc8 is generated by the motor torque mt obtained during the execution of the positioning D8 shown.
[0106] exist Figure 6 , Figure 7 In the example of , the multiple processing times for generating the state feature sc8 based on the state feature sc1 do not overlap with each other, but the present embodiment is not limited to the case where the times for generating the state feature sc do not overlap with each other. For example, the relationship between the time ts1 and the time ts2 can be any of before, after, or at the same time, and can be freely selected.
[0107] As described above, multiple processing times may overlap with each other. For example, the time for executing sampling, i.e., the processing time, may be regarded as a window (window), and the starting time of the processing time may be staggered by the time interval of sampling performed at equal intervals, i.e., the sampling time, and the window may be sampled and set as the state feature sc. For example, when the sampling time is set to 1 millisecond and the number of variables N of the state feature sc of the sampling number of a window is set to 100, the calculated state feature sc becomes a situation in which 99% of the sampling numbers are repeated and 1% are different. As described above, the method of implementing sampling by staggering windows with a predetermined sampling number in sequence is called a sliding window method. This sliding window method may be used.
[0108] exist Figure 6 , Figure 7 In the example of , the state signal ss obtained in time series is sampled with a predetermined number of samples and output as a set of state feature quantities sc. Figure 6 , Figure 7 Different methods of generating state feature quantities sc are illustrated. For example, based on the conditional signal cs of the time series, multiple statistics are calculated to form a set of state feature quantities sc or conditional feature quantities cc. The numerical value obtained by applying a statistical algorithm to the time series signal detected as the state quantity sa can be used as a statistic. In addition, it can also be a value obtained by summarizing its characteristics from a certain number of sample data. Here, there are multiple methods for statistical algorithms and general methods. Examples of statistics include average, standard deviation, dispersion, root mean square, maximum value, minimum value, peak value, wave height ratio, kurtosis, skewness, etc.
[0109] In addition, frequency analysis can also be performed on the time series signal to set it as a state feature sc or a conditional feature cc. For example, the gain of a specific frequency, the phase of a specific frequency, etc. obtained by frequency analysis can also be measured to set it as a state feature sc or a conditional feature cc. When performing processing such as statistical calculation and frequency analysis when calculating the state feature sc or the conditional feature cc, the number of samples between the processing times of the signal used in the calculation of the feature does not necessarily need to be set to the same. On the other hand, the number of variables of the calculated state feature sc or conditional feature cc is preferably set to the same.
[0110] The state feature quantity generating unit 12 arranges the state signal ss obtained in time series with a predetermined number of samples and outputs it as a set of state feature quantities sc. In the above case, the state feature quantities sc1, sc2, sc3, sc4, and sc8 contain information related to the state of the mechanical device 2 according to the respective operating conditions oc, and become a vector composed of multiple variables.
[0111] For example, the speed of the servo motor 204 is different in the state feature sc1 and the state feature sc2. In addition, the state feature sc3 includes the time when the speed of the servo motor 204 is not constant and the servo motor 204 accelerates. In addition, the positioning D4 includes the time when the servo motor 204 accelerates, the time when the speed is constant, and the time when the servo motor 204 decelerates, and the state feature sc4 samples the acceleration time. In addition, the moving distance of the servo motor 204 in the state feature sc8 is small and the maximum speed is also small.
[0112] As described above, in the mechanical device 2 used under various operating conditions oC, it is difficult to implement learning in advance for all combinations of operating conditions oC. It is also possible to adopt a method of extracting data from the time when the servo motor 204 is operated at a speed included in a predetermined speed range, or the time when the servo motor 204 is operated at an acceleration included in a predetermined acceleration range. In this method, stable data can be obtained, but there is a problem that the processing of the extracted signal is cumbersome. In addition, depending on the operating conditions oC, there may be a situation where an abnormality cannot be detected.
[0113] Furthermore, as far as the inventors know, there is no method for quantitatively evaluating the degree of deviation between the operating conditions of the detection object measured as a plurality of variables obtained in a time series and the operating conditions during learning measured as a plurality of variables obtained in a time series as described above. In addition, the problem that it is difficult to quantitatively evaluate the deviation between the above two conditions has not been recognized. Here, in Figure 6 , Figure 7In the example of , the operating condition to be detected is represented as a time series detection condition signal dcs. In addition, the operating condition at the initial learning time is represented as an initial learning condition signal lcs.
[0114] The condition feature quantity generating unit 16 generates the initial learning condition feature quantity lcc based on the initial learning condition signal lcs. In addition, the condition feature quantity dcc for detection is generated based on the detection condition signal dcs. Figure 6 In the example, Figure 6 (a) Figure 7 The command speed ds shown in (a) is set as the condition signal cs. Then, the condition feature quantity generating unit 16 sets the command speed ds included in each of the plurality of processing times as the detection condition feature quantity dcc.
[0115] The conditional feature quantity generating unit 16 sets the time series signal of a set of command speed ds from the time ts1 to the time te1 when positioning D1 is performed as the conditional feature quantity cc1. In addition, a set of command speed ds from the time ts2 to the time te2 when positioning D2 is performed is set as the conditional feature quantity cc2. In addition, a set of command speed ds from the time ts3 to the time te3 when positioning D3 is performed is set as the conditional feature quantity cc3. In addition, a set of command speed ds from the time ts4 to the time te4 when positioning D4 is performed is set as the conditional feature quantity cc4 (in Figure 6 In addition, a set of command speeds ds from the time ts8 to the time te8 when positioning D8 is performed is set as conditional feature cc8. Here, conditional feature cc1, conditional feature cc2, conditional feature cc3, conditional feature cc4, and conditional feature cc8 can each be used as a detection conditional feature dcc or as an initial learning conditional feature lcc.
[0116] As described above, the detection condition feature dcc1, detection condition feature dcc2, detection condition feature dcc3, detection condition feature dcc4, and detection condition feature dcc8 generated by the condition feature generation unit 16 correspond to the detection state feature dsc1, detection state feature dsc2, detection state feature dsc3, detection state feature dsc4, and detection state feature dsc8, respectively. Here, the detection condition feature dcc corresponds to the detection state feature dsc, which means that the state signal ss and the operating condition oc used for their respective generation are detected during the same processing time.
[0117] In addition, Figure 6 and Figure 7In the example, for ease of understanding, the sampling frequencies of the command speed ds and the motor torque mt are set to be the same. In addition, the number of variables of a set of state feature quantities sc and the number of variables of a set of conditional feature quantities cc are set to be the same. However, the present embodiment is not limited to the above-described method. For example, the sampling frequency of the conditional signal cs and the sampling frequency of the state signal ss may be different. In addition, for example, the number of variables of a set of state feature quantities sc and the number of variables of a set of conditional feature quantities cc may be different from each other. In addition, if the detection state signal dss and the detection condition signal dcs used for the calculation of the corresponding detection state feature quantity dsc and the detection conditional feature quantity dcc are obtained during the same processing time, the abnormality degree an and the unknown degree un can be calculated based on the data obtained at the same processing time, so that abnormality detection can be performed more accurately.
[0118] The method by which the conditional feature quantity generating unit 16 generates the conditional feature quantity cc may be the same as or different from the method by which the state feature quantity generating unit 12 generates the state feature quantity sc. For example, the method for generating the conditional feature quantity cc may use statistical calculation, and the method for generating the state feature quantity sc may use frequency analysis.
[0119] The initial state learning unit 13 performs learning based on the initial learning state feature quantity lsc and outputs the learning result as the initial state learning result slr. In addition, the initial condition learning unit 17 performs learning based on the initial learning condition feature quantity lcc and outputs the learning result as the initial condition learning result clr. Figure 8 This is a diagram showing an example of an autoencoder according to this embodiment.
[0120] Autoencoder is a type of neural network model. Figure 8 The autoencoder illustrated in has an input layer consisting of nodes X1 to X3, an intermediate layer consisting of nodes Y1 and Y2, and an output layer consisting of nodes Z1 to Z3. Figure 8 The nodes of the autoencoder are connected at multiple edges given weights W11 to W26 as parameters. Figure 8 The autoencoder passes the value calculated by the function set at the edge to the node of the next layer connected via the edge based on the value input to the input layer, and finally obtains the result from the output layer.
[0121] That is, the neural network as a whole constitutes a complex function. The autoencoder is a type of unsupervised learner that learns each parameter (weight) in such a way that the data input to the input layer and the data output from the output layer are close. Multiple input data are prepared, and the parameters as weights are adjusted for any data to reduce the error between input and output, thereby the network continuously learns the characteristics of the input signal.
[0122] exist Figure 8 In the example, for ease of understanding, the number of nodes in the input layer is set to 3, and the number of nodes in the output layer is set to 3, but the number of input and output nodes does not need to be the same as the number of input variables N set as the object (for example, in Figure 6 In the example, N=100) matches the number of input and output nodes. Figure 6 , 7 In the example of , they can be 100 respectively. In addition, for the purpose of explanation, Figure 8 In the example, the number of nodes in the middle layer is set to 2, and the total number of the middle layers is set to 1, but the number of nodes in the middle layer and the number of layers are not limited to this.
[0123] The initial state learning unit 13 learns the neural network in such a way that if the state feature sc1 as the initial learning state feature lsc is input to the network, an estimated value of the state feature sc1, i.e., sc1', is obtained as an output. And the learned neural network model is output as the initial state learning result slr. The initial state learning result slr can be any one that can identify the model, for example, it can be the parameters of the model, the structure of the model, etc. The initial condition learning unit 17 learns the neural network in such a way that if the condition feature cc1 as the initial learning condition feature lcc is input to the network, an estimated value of the condition feature cc1, i.e., cc1', is obtained as an output. The learned neural network model is output as the initial condition learning result clr. The initial condition learning result clr can be any one that can identify the model, for example, it can be the parameters of the model, the structure of the model, etc.
[0124] exist Figure 8In the embodiment, as an example of the initial state learning unit 13 and the initial condition learning unit 17, the structure using the autoencoder is described, but the present embodiment is not limited to this structure. As examples of the structure of the initial state learning unit 13 and the initial condition learning unit 17 different from the autoencoder, there can be cited self-organizing map (Self-organizing map, SOM), MT method (Mahalanobis Taguchi Method), principal component analysis (Principal component analysis, PCA), one class support vector machine (One class support vector machine, OCSVM), k nearest neighbor method (kNearest Neighbors method), isolation forest (Isolation Forest), etc. In a method different from the above-mentioned exemplified method, if it is a learning method that learns the input feature quantity in advance and can evaluate the degree of deviation between the characteristics of the feature quantity obtained after learning and the characteristics of the trained feature quantity, it can also be used as a learning method for the initial state learning unit 13 or the initial condition learning unit 17.
[0125] exist Figure 8 In the example, the initial state learning unit 13 and the initial condition learning unit 17 are assumed to have the same structure and are described. The structures of the initial state learning unit 13 and the initial condition learning unit 17 do not necessarily have to be the same, and different learning methods may be combined. However, in the case where the state feature quantity sc and the condition feature quantity cc maintain the same complexity, it is preferred that the initial state learning unit 13 and the initial condition learning unit 17 are composed of learning methods with the same degree of explanation ability. For example, it is preferred that the number of nodes and the number of edges of the neural network of the initial state learning unit 13 and the initial condition learning unit 17 are the same. In addition, it is preferred that the number of input data of the initial state learning unit 13 and the initial condition learning unit 17 be the same.
[0126] The operation of the abnormality calculation unit 14 and the unknown degree calculation unit 18 is described. The abnormality calculation unit 14 calculates the abnormality an based on the initial state learning result slr and the detection state signal dss. The abnormality an can be set as the degree of deviation between the initial learning state signal lss and the detection state signal dss. For example, the abnormality calculation unit 14 inputs the detection state feature quantity dsc to the model constructed using the initial state learning result slr. Moreover, the degree of deviation between the initial learning state feature quantity lsc and the detection state feature quantity dsc can be calculated, and the degree of deviation, information indicating the degree of deviation, etc. can be used as the abnormality an.
[0127] like Figure 8As shown in the example, when learning using an autoencoder, for example, the difference between the state feature quantity sc input to the autoencoder and the estimated value of the state feature quantity sc output from the autoencoder can be used as the abnormality an. The state feature quantity sc and the estimated value of the state feature quantity sc are both vectors, so the difference between the two is also a vector. Figure 8 In the example of , the abnormality an is evaluated as a scalar value, so the difference between the two, that is, the square root of the sum of the squares of the residual vectors is set as the abnormality an.
[0128] As a method for calculating the abnormality an using an autoencoder, there is also a method in which, instead of the difference between the input and output as described above, the difference between the value of the intermediate layer during learning and the value of the intermediate layer during evaluation is set as the abnormality an. In the case where there is one node in the intermediate layer, the difference between the average value of the intermediate layer during learning and the value of the intermediate layer during inference (detection) can be calculated, and the magnitude of the difference can be used as the abnormality an. In the case where there are multiple nodes in the intermediate layer, similarly to the case of calculating the difference between the input and output described above, the square root of the sum of the squares of the residual vectors obtained by subtracting the average value of the intermediate layer during learning from the value of each intermediate layer during inference is set as the abnormality an.
[0129] In the learning method, when using self-organizing maps, the minimum quantization error (MQE) can be used as the abnormality an. When using principal component analysis, the T2 statistic and Q statistic can be used as the abnormality an. When using a one-class support vector machine, the distance from the origin in the mapped space can be used as the abnormality an. When using the k-nearest neighbor method, the distance between the feature quantity to be evaluated and the k trained feature quantities that are close to the feature quantity to be evaluated can be used as the abnormality an.
[0130] The unknown degree calculation unit 18 calculates the degree of deviation between the initial learning condition signal lcs and the detection condition signal dcs, that is, the unknown degree un, based on the initial condition learning result clr and the detection condition signal dcs. For example, the unknown degree calculation unit 18 constructs a model using the initial condition learning result clr. Then, the detection condition feature quantity dcc is input to the constructed model, and the degree of deviation between the output of the constructed model and the detection condition feature quantity dcc is calculated. The degree of deviation can be regarded as the unknown degree un.
[0131] like Figure 8As shown in the example of , when the autoencoder is used for learning, the difference between the input detection condition feature dcc and the estimated value output from the model can be used as the unknown degree un. The detection condition feature dcc and the estimated value of the detection condition feature dcc output from the model are both vectors, so the difference between the vectors also becomes a vector. Therefore, the unknown degree un is evaluated as a scalar value, so the square root of the sum of the squares of the residual vectors, which is the difference between the vectors, can be used as the unknown degree un.
[0132] Fig. 9 1 is a diagram showing an example of a structure in which the abnormality determination unit 19 is omitted from the abnormality detection device 1 according to the present embodiment. The structure in which the abnormality determination unit 19 is omitted from the structure of the abnormality detection device 1 is referred to as an abnormality detection device 1p. Fig. 9 In, with Figure 1 The same structural elements and signals are marked with the same reference numerals.
[0133] The abnormality detection device 1p is Figure 1 The structure of the abnormality detection device 1 shown in the figure omits four structural elements, namely, the condition signal generating unit 15, the condition feature quantity generating unit 16, the initial condition learning unit 17, and the unknown degree calculating unit 18. Figure 1 The abnormality determination unit 19 of the abnormality detection device 1 shown in the figure has an abnormality determination unit 19a. The difference between the abnormality determination unit 19a and the abnormality determination unit 19 is described. The abnormality determination unit 19 performs determination based on the abnormality degree an and the unknown degree un. On the other hand, the abnormality determination unit 19a does not use the unknown degree un, but performs determination based on the abnormality degree an.
[0134] The abnormality detection device 1p is the same as the abnormality detection device 1 in terms of the points other than those described above. Fig.10 1 is an example of the time change of the abnormality degree an and the determination result jr generated by the configuration in which the abnormality determination unit 19 is omitted from the abnormality detection device 1 according to the present embodiment. Fig.10 This is an example of the time change of the abnormality degree an and the determination result jr generated by the abnormality detection device 1p. Fig.11 The abnormality degree an, the unknown degree un, and the time variation of the determination result jr generated by the configuration in which the abnormality determination unit 19 is omitted from the abnormality detection device 1 according to the present embodiment are compared. Fig.10 Another example of a different Fig.11 This is an example of the time change of the abnormality degree an and the determination result jr generated by the abnormality detection device 1p.
[0135] Moreover, about Fig.10 , Fig.11The abnormality detection device 1p outputs the determination result jr by the abnormality determination unit 19a. Fig.12 This is an operation in which the abnormality determination unit 19 of the abnormality detection device 1 performs abnormality determination using the unknown degree un. The effects of using the unknown degree un will be described by comparing the two.
[0136] Fig.10 , Fig.11 The horizontal axis represents time, and the unit is time (hr). Fig.10 (a) and Fig.11 (a) is the time variation of the abnormality an. Fig.10 (a) and Fig.11 The vertical axis of (a) represents the abnormality degree an. Fig.10 (b) and Fig.11 (b) is the time change of the determination result jr. Fig.10 (b) and Fig.11 The vertical axis of (b) represents the judgment result jr. Fig.10 (a) and Fig.10 In (b), two positions with the same number on the time axis represent the same time. Fig.11 (a) and Fig.11 In (b), two positions with the same number on the time axis represent the same time.
[0137] Fig.10 The example of is to plot the value of the abnormality an from the moment ta1 to the moment tg1, which is the time ta1 elapsed from the start of the operation of the abnormality detection device 1p, in units of one hour. Between the time ta1 and the moment ta1, the completion of the initial state learning of the mechanical device 2 is completed. Between the moment ta1 and the moment td1, although the abnormality an contains a certain amount of fluctuation, it is all plotted between 0 and 1, roughly showing the characteristic of small time change. In addition, between the moment td1 and the moment tg1, the abnormality an gradually increases with the passage of operating time. This reflects that after the moment td1, the mechanical device 2 gradually deteriorates. The abnormality an exceeds 1 for the first time at the moment te1, and after the moment tf1, the abnormality an exceeds 1.
[0138] exist Fig.10 In the example of (a), the threshold value THF1 of the abnormality degree an used by the abnormality determination unit 19a when calculating the determination result jr is set to 1. The abnormality determination unit 19a determines that it is abnormal when the value of the abnormality degree an exceeds the threshold value THF1. Moreover, when the value of the abnormality degree an is the same as the threshold value THF1 or is lower than the threshold value THF1, it determines that it is normal.
[0139] And, in Fig.10 In (a), the true abnormality TRUE1 of the mechanical device 2 is illustrated by a thick solid line. The true abnormality TRUE1 is the abnormality an based on the state quantity sa after the influence of the interference is completely eliminated, and is a hypothetical abnormality an for easy understanding of the description. At each time point, the state quantity sa for detection is completely eliminated and the influence of the interference is obtained. Based on the state quantity sa, the abnormality an is obtained when the state characteristic quantity dsc for detection and the abnormality an are calculated. As an example of interference, for example, during initial state learning and detection, the operating condition oc of the mechanical device 2 is changed. In this case, when the operating condition oc is restored to the initial state learning, the state quantity sa for detection is measured, thereby eliminating the interference. In reality, when it is difficult to eliminate the interference or it is impossible to eliminate the interference, here, the abnormality an calculated under the assumption that the interference can be eliminated and the true abnormality is calculated is called the true abnormality TRUE1. In addition, the true abnormality TRUE1 is different from the abnormality an estimated by the abnormality detection device 1p. In addition, the true abnormality TRUE1 is not affected by interference from operating conditions oc and the like. The ideal operation of a general abnormality detection device can be said to be aimed at detecting abnormalities by estimating the true abnormality TRUE1. In other words, when the abnormality an correctly represents the state of the mechanical device 2, the abnormality an becomes the same value as the true abnormality TRUE1. Detecting the true abnormality TRUE1 is an ideal situation, but actual abnormality detection devices are affected by interference and therefore often fail to detect the true abnormality TRUE1. Fig.10 In FIG. 1 , the true abnormality degree TRUE1 is plotted for easy understanding, but the true abnormality degree TRUE1 does not necessarily need to be calculated.
[0140] exist Fig.10 In (b), the time change of the determination result jr obtained by the abnormality detection device 1p is plotted in units of one hour. When the mechanical device 2 is normal, the determination result jr is set to 0, and when the mechanical device 2 is abnormal, the determination result jr is set to 1. The output method of the determination result jr obtained by the abnormality determination unit 19a is not limited to the above method. The output method of the determination result jr can also be able to distinguish the normal and abnormal categories based on the determination result jr. In addition, it is also possible to know the degree of abnormality.
[0141] In addition, the output method of the judgment result jr includes not only the output of a signal including the information of the judgment result jr, the display of the judgment result jr to the operator (including the case of not displaying), the issuance of an alarm (such as sounds such as sirens, red lights, etc.), the stopping of the alarm (including the case of no alarm output), etc., but also the stopping of the mechanical device 2, the reduction of the operating speed of the mechanical device 2, the stopping of a device connected to the mechanical device 2, the start-up instruction of the maintenance device of the mechanical device 2, etc.
[0142] exist Fig.10 In the example of (b), the determination result jr from time ta1 to time te1 is a normal value indicating a normal value. That is, the value of the determination result jr is 0. Moreover, the determination result jr changes from a normal value to an abnormal value at least once at time te1. The determination result jr from time te1 to time tf1 is a mixture of normal values and abnormal values due to the fluctuation of the abnormality an.
[0143] according to Fig.10 (a) The mechanical device 2 starts to deteriorate at time td1, and the deterioration gradually intensifies after time td1. The determination result jr after time td1 is an abnormal value and does not correspond to a false detection (a normal state is determined to be abnormal). Fig.10 In the case of the situation shown in FIG. 1 , the abnormality detection device 1p can detect abnormalities without causing false detection. That is, if the mechanical device 2 is in Fig.10 In the situation shown, even if the unknown degree un is not used in abnormality detection, false detection will not occur. In addition, the abnormality detection device 1p can output the abnormality degree an close to the true abnormality degree TRUE1.
[0144] right Fig.11 As mentioned above, Fig.11 This is an example of the result detected by the abnormality detection device 1p. Fig.11 The time periods shown and Fig.10 The time periods shown are different from each other. Fig.11 In the example of , at time td1', the degradation of the mechanical device 2 begins, and after time td1', the mechanical device 2 gradually degrades. Moreover, between time tb1' and time tc1', the speed of the motor 20 is changed to a value different from the speed of the motor 20 during the initial state learning. Moreover, except for the time between time tb1' and time tc1', the time between time ta1' and time tg1' is set to be the same as the speed of the motor during the learning.
[0145] Fig.11 (a) shows the temporal change of the abnormality degree an. Fig.11(b) shows the temporal change of the determination result jr. Fig.11 (a) and Fig.11 The horizontal axis of (b) represents time, and the unit is time (hr). Fig.11 (a) Timeline and Fig.11 In the time axis of (b), the times indicated by the same reference numerals are the same times. Fig.11 In the example of FIG. 1 , data from time ta1 ′ to time tg1 ′, which is the time when the operation of the abnormality detection device 1 p is started and the time ta1 ′ has elapsed, are shown. Furthermore, a case where the initial state learning of the initial state learning unit 13 is completed before time ta1 ′ is described.
[0146] exist Fig.11 In (a), the vertical axis represents the abnormality an, and the data points of the abnormality an are plotted every hour. Although the abnormality an contains a certain amount of fluctuations between time ta1' and time tb1' and between time tc1' and time td1', all of them are plotted in the range of 0 to 1, which roughly shows the characteristic of small time change.
[0147] In addition, from time td1' to time tg1', the abnormality degree an shows a characteristic that it gradually increases with the passage of operating time, although it contains a certain amount of fluctuation. Fig.10 (a) and Fig.11 (a) is used to explain the difference. Fig.11 In (a), the abnormality an exceeds 1 between time tb1' and time tc1'. Fig.10 In (a), the abnormality degree an is maintained at a value lower than 1 between the time ta1 and the time td1. Fig.11 The increase in the abnormality an from time tb1' to time tc1' in (a) is not caused by the deterioration of the mechanical device 2. The increase in the abnormality an is caused by the change of the operating condition from time tb1' to time tc1' to the initial learning, that is, the change of the operating condition oc.
[0148] like Fig.11 As shown in (a), the threshold value THF1' related to the abnormality an is set to 1. Moreover, the threshold value THF1' is determined based on the distribution of the abnormality an when the mechanical device 2 is normal. Fig.11 In (a), the true abnormality TRUE1' of the mechanical device 2 is indicated by a thick solid line. Fig.11 The true abnormality TRUE1' is related to the description of Fig.10 The description related to the true abnormality TRUE1 recorded in the description is the same as that recorded in the description, so it is omitted.
[0149] in addition, Fig.11 (b) The determination result jr is plotted on the vertical axis. Fig.11 The judgment result jr of (b) is shown by connecting the data points obtained every hour with lines. Fig.11 In (b), as an example of the form of the determination result jr, when the mechanical device 2 is normal, the value of the determination result jr is plotted as 0, and when the mechanical device 2 is abnormal, the value of the determination result jr is plotted as 1.
[0150] The abnormality determination unit 19a Fig.10 The abnormality an of (a) is compared with the threshold value THF1'. When the abnormality an is less than or equal to the threshold value THF1', the state of the mechanical device 2 is determined to be normal, and the determination result jr is set to 0. On the other hand, when the abnormality an exceeds the threshold value THF1', the state of the mechanical device 2 is regarded as abnormal, and the determination result jr is set to 1. Here, the abnormality an is less than or equal to the threshold value THF1' means that the value of the abnormality an is lower than the threshold value THF1', or the value of the abnormality an is the same as the threshold value THF1'.
[0151] like Fig.11 As shown in (b), the abnormality determination unit 19a outputs 0 as a value indicating normality as the determination result jr for the two time periods from the time ta1' to the time tb1' and from the time tc1' to the time te1'. In addition, since the abnormality degree an exceeds 1 between the time tb1' and the time tc1', the abnormality determination unit 19a outputs a value 1 indicating abnormality as the determination result jr.
[0152] right Fig.10 , Fig.11 For comparison, as described in the description, Fig.10 In the example of , the operating condition oc during detection is maintained the same as that during initial learning. Fig.11 In the example of , during a portion of the detection time, the operating condition oc is changed to the initial state learning time. Fig.10 In the example of , no false detection occurs in the detection result of the abnormality detection device 1p. Fig.11 In the example of , due to the change of the operating condition oc from the initial learning, an erroneous detection occurs in the detection result of the abnormality detection device 1p. Here, the event that the mechanical device 2 outputs a determination result that the mechanical device 2 is abnormal when the mechanical device 2 is normal is called an erroneous detection.
[0153] according to Fig.10 , Fig.11In contrast, when different time periods of the operating condition οc during initial state learning and detection are included in the detection time, there is a possibility that false detection may occur in the abnormality detection device 1p that outputs the judgment result jr without using the unknown degree un.
[0154] use Fig.12 , the operation of the abnormality determination unit 19 of the abnormality detection device 1 according to the present embodiment will be described. Fig.12 : is a diagram showing the time changes of the abnormality degree an, the unknown degree un, and the determination result jr generated by the abnormality detection device 1 according to the present embodiment. Fig.12 , the time change when the mechanical device 2 gradually deteriorates after the time td1″ is shown. Fig.12 From time tb1" to time tc1", the speed of the motor 20 is changed to a setting different from that during the initial learning. Fig.12 During the period from time ta1 ″ to time tb1 ″ and during the period from time tc1 ″ to time tg1 ″, the speed of the motor 20 is set to the same setting as that during the initial learning.
[0155] Fig.12 (a) shows the temporal change of the abnormality degree an. Fig.12 (b) shows the time variation of the unknown degree un. Fig.12 (c) shows the temporal change of the determination result jr. Fig.12 (a) Fig.12 (b) and Fig.12 The horizontal axis of (c) represents time, and the unit is time (hr). Fig.12 (a) Fig.12 (b) Fig.12 The times on the time axis of (c) with the same reference numerals represent the same time. Fig.12 In the example, data from time ta1″ to time tg1″ after a time ta1″ has passed since the start of the operation of the abnormality detection device 1 is shown. In addition, it is assumed that the initial learning related to the mechanical device 2, that is, the initial state learning and the initial condition learning are completed before the time ta1″.
[0156] Fig.12 The abnormality an of (a) is plotted as data points per hour. Fig.12 In (a), although the value of the abnormality an from time ta1" to time tb1" and from time tc1" to time td1" contains a certain amount of fluctuation, all of them are within the range of 0 to 1, and the change in value is small. In addition, the abnormality an gradually increases with the passage of time from time td1" to time tg1".
[0157] In addition, Fig.12 In (a), the abnormality an exceeds 1 from time tb1" to time tc1". The increase in the abnormality an in this period is related to Fig.11 The contents of the description of the abnormality degree an between time tb1' and time tc1' in (a) are the same. Fig.12 The increase in the abnormality an from time tb1" to time tc1" in (a) is not caused by the increase in the abnormality an due to the deterioration of the mechanical device 2, but is caused by the operating conditions, in other words, the change in the operating condition oc. Fig.11 Similarly, from time tb1' to time tc1', it is Fig.12 This is caused by the fact that the speed of the motor 20 changes from that during the initial learning from the time tb1″ to the time tc1″.
[0158] exist Fig.12 In the case of the mechanical device 2, no degradation occurs between the time tb1" and the time tc1". As described above, although it is not an abnormal state, the characteristics of the state feature sc obtained between the time tb1" and the time tc1" are different from the characteristics of the state feature sc at the time of initial state learning. Due to the change of the state feature sc from the initial learning, Fig.11 The abnormality degree an shown in (a) becomes a large value exceeding THF1".
[0159] in addition, Fig.12 The abnormality an of becomes a value exceeding 1 at time te1”, and all the abnormality an exceeds 1 after time tf1”. Fig.12 As shown in (a), the threshold value THF1" related to the abnormality is set to 1. In addition, the threshold value THF1" can be determined based on the distribution of the abnormality an when the mechanical device 2 is normal. Fig.12 In (a), unlike the abnormality an output by the abnormality detection device 1, the true abnormality TRUE1" of the mechanical device 2 which will not be affected by the disturbance of the operating condition oc is represented by a thick solid line. Fig.10 The description of the true abnormality is the same as that of TRUE1.
[0160] Fig.12 The vertical axis of (b) is the unknown degree un calculated by the unknown degree calculation unit 18. Fig.12In (b), the unknown degree un is plotted between 0 and 1 from time ta1" to time tb1" and from time tc1" to time tg1". On the other hand, the unknown degree un is plotted between 1 and 2 from time tb1" to time tc1". Fig.12 As shown in (b), the threshold THU1" related to the unknown degree un is set to 1.
[0161] The threshold value THU1” can be determined based on the distribution of the unknown degree un when the initial condition learning is implemented. The threshold value THU1” can be determined based on the distribution of the unknown degree un when the initial condition learning is implemented, thereby being able to determine whether the abnormality an accurately represents the state of the mechanical device 2 based on the unknown degree un. For example, when the deviation between the unknown degree un calculated by the abnormality determination unit 19 during the detection time and the distribution of the unknown degree un during the initial condition learning is large, it can be determined that the abnormality an does not accurately represent the state of the mechanical device 2 based on the unknown degree un. In addition, for example, when the deviation between the unknown degree un calculated by the abnormality determination unit 19 and the distribution of the unknown degree un during the initial condition learning is small, it can be determined that the abnormality an accurately represents the state of the mechanical device 2 based on the unknown degree un.
[0162] in addition, Fig.12 The vertical axis of (c) represents the determination result jr of the presence or absence of abnormality in the mechanical device 2 determined by the abnormality determination unit 19. The determination result jr is Fig.12 In the example of (c), the data points calculated every hour are connected by lines and plotted. As an example of the output method of the determination result jr, when the state of the mechanical device 2 is normal, the determination result jr is set to 0, and when the state of the mechanical device 2 is abnormal, the determination result jr is set to 1. The output method of the determination result jr of this embodiment is not limited to the above method.
[0163] The abnormality determination unit 19 Fig.12 (a) shows the abnormality an and the threshold value THF1". Fig.11 The unknown degree un shown in (b) is compared with the threshold value THU1". Then, when the abnormality degree an exceeds the threshold value THF1" and the unknown degree un is less than or equal to the threshold value THU1", the abnormality determination unit 19 regards the state of the mechanical device 2 as abnormal and sets the determination result jr to 1.
[0164] In cases other than those described above, the state of the mechanical device 2 is considered normal, and the judgment result jr is set to 0. Here, the cases other than those described above are at least one of the first case or the second case described below. The first case is a case where the abnormality an is less than or equal to the threshold THF1". The second state is a case where the unknown degree un exceeds the threshold THU1".
[0165] The abnormality determination unit 19 uses the unknown degree un when outputting the determination result jr. Fig.12 As shown in (c), the abnormality determination unit 19 outputs a value 0 indicating normality as the determination result jr between the time ta1" and the time te1". Furthermore, no false detection occurs in the abnormality detection device 1. On the other hand, Fig.11 The abnormality determination unit 19a described in the above does not use the unknown degree un. Fig.11 As shown in (b), between time tb1' and time tc1', a value 1 indicating an abnormality is generated as the determination result jr. Furthermore, an erroneous detection occurs in the abnormality detection device 1a. Fig.12 and Fig.11 As a comparative description, the abnormality detection device 1 of the present embodiment can perform abnormality detection with less false detection by utilizing the unknown degree un.
[0166] In addition, as shown above, instead of setting threshold values for the abnormality an and the unknown degree un, the structure of the abnormality determination unit 19 can be replaced by using a value obtained by dividing the abnormality an by the unknown degree un (a value obtained by dividing the abnormality an by the unknown degree un. That is, an / un), etc., to reflect the difference in the operating condition oc in the output of the determination result jr, thereby constituting an abnormality detection device 1 that is not prone to false detection. For example, a threshold value is set for the value obtained by dividing the abnormality an by the unknown degree un. And, when the value exceeds the threshold value, it is determined to be abnormal. In addition, when the value obtained by dividing the abnormality an by the unknown degree un is less than or equal to the threshold value, it is determined to be normal.
[0167] Fig.131 is a diagram showing an example of the action flow of the abnormality determination unit 19 involved in the present embodiment. In step S101, the abnormality determination unit 19 obtains a group of the abnormality an calculated by the abnormality calculation unit 14 and the unknown degree un calculated by the unknown degree calculation unit 18. A group of abnormality an and unknown degree un preferably corresponds to each other. In other words, the abnormality an and the unknown degree un are preferably based on information obtained between the same detection time. That is, the detection state signal dss used for the generation of the abnormality an and the detection condition signal dcs used for the generation of the unknown degree un are preferably obtained at the same detection time. In addition, for example, when the unknown degree un is calculated by using the command speed ds when the motor 20 is running as the detection condition signal dcs, the abnormality an is determined by using the torque when the motor 20 is running by the command speed ds as the detection state signal dss. As described above, it is preferred that the abnormality an and the unknown degree un correspond to each other.
[0168] The detection condition signal dcs used to generate the unknown degree un corresponds to the detection state signal dss used to generate the abnormality degree an, thereby enabling high-precision abnormality detection. In addition, abnormality detection with less occurrence of false detection and omission can be performed.
[0169] Next, in step S102, the abnormality determination unit 19 proceeds to step S103 when the abnormality degree an exceeds the threshold value THF1” and the unknown degree un is less than or equal to the threshold value THU1”. In other cases, it proceeds to step S104. Here, the case where the unknown degree un is less than or equal to the threshold value THU1” is either the case where the unknown degree un is the same as the threshold value THU1” or the case where the unknown degree un is less than the threshold value THU1”. Step S102 becomes a step in which the abnormality determination unit 19 implements a determination of whether it is abnormal or normal.
[0170] In step S103, the abnormality determination unit 19 outputs the determination result indicating that an abnormality has occurred in the mechanical device 2 as the determination result jr. In step S104, the abnormality determination unit 19 outputs the determination result indicating that the mechanical device 2 is normal as the determination result jr. Steps S103 and S104 may include an action of notifying the user of the determination result jr through an interface or the like as needed. Fig.12 Description of the action flow.
[0171] Fig.14 1 is a block diagram showing an example of the structure of the mechanical system 100x according to the present embodiment. Fig.14The following mainly describes the differences from the abnormality detection device 1. The mechanical system 100x includes an abnormality detection device 1x. The abnormality detection device 1x is different from the abnormality detection device 1 in that the learning model is made universal when monitoring the mechanical device 2-n from the plurality of mechanical devices 2-1. When the plurality of mechanical devices 2-1 to 2-n have substantially the same characteristics, the learning model of the present embodiment is used. Fig.14 The effect of the abnormality detection device 1x described above is more greatly exerted.
[0172] As an example of the above-mentioned equivalent characteristics, for example, the mechanical devices 2-1 to 2-n are manufactured with the same specifications and operate under different operating conditions oc. In addition, as another example, the following situation can be cited, that is, the motor that drives the mechanical devices 2-1 to 2-n is common, and the abnormality caused by the operation of the motor is specifically detected, and the operating condition oc and the state quantity sa are related to the motor or the driven object of the motor. The abnormality detection device 1x obtains the state signal ss-k from the mechanical device 2-k (k is an integer from 1 to n). In addition, the operating condition oc-k is obtained from the control device 3-k.
[0173] The initial state learning unit 13x outputs the initial state learning result slr-1 based on the initial learning state feature quantity lsc-1. Furthermore, the abnormality calculation unit 14x outputs the abnormality an-k related to the mechanical device 2-k based on the initial state learning result slr-1 and the detection state feature quantity dsc-k.
[0174] In addition, the initial condition learning unit 17x outputs the initial condition learning result clr-1 based on the initial learning condition feature quantity lcc-1. Furthermore, the unknown degree calculation unit 18x outputs the unknown degree un-k related to the mechanical device 2-k based on the initial condition learning result clr-1 and the detection condition feature quantity dcc-k.
[0175] The abnormality determination unit 19 outputs the determination result jr-k related to the mechanical device 2-k based on the abnormality degree an-k and the unknown degree un-k. The abnormality detection device 1x makes the learning model of the mechanical devices 2-k from k=1 to n common, so that the calculation load can be reduced compared with the case where the calculation model is prepared for each mechanical device 2-k from k=1 to n. In addition, multiple data for initial state learning can be prepared in parallel. The above is related to Fig.14 Description of variations of the present embodiment shown.
[0176] right Figure 1 A variation of the abnormality detection device 1 of the present embodiment shown in FIG. Fig.13In the flowchart, one threshold is set for each of the abnormality an and the unknown degree un. Multiple thresholds can be set for the values of the abnormality an and the unknown degree un to separate them into multiple ranges, and based on the combination of the range containing the abnormality an and the range containing the unknown degree un, it is determined whether it is abnormal or normal.
[0177] In addition, Fig.13 In the flowchart, either abnormality or normality is output as the judgment result jr, but the abnormality judgment unit 19 can also be configured to output three or more judgment results jr corresponding to the abnormality degree an and the unknown degree un. For example, it can be configured to determine the judgment results jr of four stages: severe abnormality, mild abnormality, mild normality that requires further observation, and normality that does not require observation. Moreover, in this case, it can be configured to set multiple thresholds for the abnormality degree an or the unknown degree un. In addition, in Fig.13 In the flowchart of , one set of abnormality degree an and unknown degree un is used to perform abnormality detection, but multiple sets of abnormality degree an and unknown degree un can also be calculated, and the determination result of abnormal state or normal state can be output for each set. Here, different abnormality degrees an can be used according to the group, or the same abnormality degree an can be used. In addition, different unknown degrees un can be used according to the group, or the same unknown degree un can be used. In addition, the above-described modified examples can also be combined with each other and implemented.
[0178] The control device of Patent Document 1 uses a single numerical value, in other words, a single scalar value to quantitatively express the load condition. In the control device of Patent Document 1, when the load condition cannot be expressed by a single numerical value, the value of the second threshold value becomes inaccurate, and false detection, omission, etc. are also prone to occur in the judgment result. As examples of situations where the change in the state of the mechanical device cannot be expressed by the load condition, there are situations where the state of the mechanical equipment changes complexly over time, the load conditions of the mechanical equipment are the same, and the external environment changes, etc. According to the abnormality detection device 1 of this embodiment, conditional learning using time series signals is performed. Therefore, even if complex changes occur between the time when the learning model is generated and the time when it is evaluated (detected), accurate judgment can be made.
[0179] An example of an abnormality detection device 1 described in this embodiment includes a state signal generating unit 11, a condition signal generating unit 15, a state feature quantity generating unit 12, a condition feature quantity generating unit 16, an initial state learning unit 13, an initial condition learning unit 17, an abnormality degree calculating unit 14 and an unknown degree calculating unit 18.
[0180] The state signal generating unit 11 generates a state signal ss obtained by detecting the state of the mechanical device 2 in a time series. The condition signal generating unit 15 generates a condition signal cs obtained by detecting the operating condition indicating the operating condition of the mechanical device 2 in a time series. The state feature quantity generating unit 12 generates a state feature quantity sc based on the state signal ss. The condition feature quantity generating unit 16 generates a condition feature quantity cc based on the condition signal cs. The initial state learning unit 13 outputs the result of learning based on the state feature quantity sc during initial state learning, that is, the initial learning state feature quantity lsc as an initial state learning result slr.
[0181] The initial condition learning unit 17 outputs the result of learning based on the condition feature quantity cc during initial condition learning, i.e., the condition feature quantity lcc for initial learning, as the initial condition learning result clr. The abnormality degree calculation unit 14 obtains the initial state learning result slr or the additional state learning result aslr as the state learning result, and calculates the abnormality degree an based on the state learning result and the state feature quantity sc during detection, i.e., the state feature quantity dsc for detection. The unknown degree calculation unit 18 obtains the initial condition learning result clr or the additional condition learning result aclr as the condition learning result, and calculates the unknown degree un based on the condition learning result and the condition feature quantity cc during detection, i.e., the condition feature quantity dcc for detection. Here, it is preferred that the detection time of the condition feature quantity cc is the same as the detection time of the state feature quantity sc.
[0182] In addition, the abnormality detection device 1 of the present embodiment may include an abnormality determination unit 19. The abnormality determination unit 19 detects the abnormality of the mechanical device 2 based on the abnormality degree an and the unknown degree un. In addition, the abnormality determination unit 19 may determine that the state of the mechanical device 2 is abnormal when the abnormality degree an is greater than a predetermined first threshold value and the unknown degree un is less than a predetermined second threshold value. In addition, when the abnormality degree an is less than or equal to the first threshold value or the unknown degree un is greater than or equal to the second threshold value, the state of the mechanical device 2 is determined to be normal.
[0183] In addition, the conditional feature quantity generating unit 16 can generate multiple statistics calculated from the conditional signal cs at multiple time points as the conditional feature quantity cc. In addition, the conditional feature quantity generating unit 16 can generate a frequency characteristic obtained by frequency analysis of the conditional signal cs of the time series as the conditional feature quantity cc. In addition, the mechanical device 2 can be driven by the motor 20 to operate. Moreover, the operating condition oc can be a control signal that specifies the shape of the time response of at least any one of the position of the motor 20, the speed of the motor 20, the acceleration of the motor 20, the jump of the motor 20, and the driving force of the motor 20.
[0184] An example of a mechanical system described in this embodiment includes a mechanical device 2, a state signal generating unit 11, a condition signal generating unit 15, a state feature quantity generating unit 12, a condition feature quantity generating unit 16, an initial state learning unit 13, an initial condition learning unit 17, an abnormality degree calculation unit 14 and an unknown degree calculation unit 18.
[0185] The state signal generating unit 11 generates a state signal ss obtained by detecting the state of the mechanical device 2 in a time series. The condition signal generating unit 15 generates a condition signal cs obtained by detecting the operating condition indicating the operating condition of the mechanical device 2 in a time series. The state feature quantity generating unit 12 generates a state feature quantity sc based on the state signal ss. The condition feature quantity generating unit 16 generates a condition feature quantity cc based on the condition signal cs. The initial state learning unit 13 outputs the result of learning based on the state feature quantity sc during initial state learning, that is, the initial learning state feature quantity lsc as an initial state learning result slr.
[0186] The initial condition learning unit 17 outputs the result of learning based on the condition feature quantity cc during initial condition learning, i.e., the condition feature quantity lcc for initial learning, as the initial condition learning result clr. The abnormality degree calculation unit 14 obtains the initial state learning result slr or the additional state learning result aslr as the state learning result, and calculates the abnormality degree an based on the state learning result and the state feature quantity sc during detection, i.e., the state feature quantity dsc for detection. The unknown degree calculation unit 18 obtains the initial condition learning result clr or the additional condition learning result aclr as the condition learning result, and calculates the unknown degree un based on the condition learning result and the condition feature quantity cc during detection, i.e., the condition feature quantity dcc for detection. Here, it is preferred that the detection time of the condition feature quantity cc is the same as the detection time of the state feature quantity sc.
[0187] An example of an abnormality detection method described in this embodiment includes a state signal generating step, a condition signal generating step, a state feature quantity generating step, a condition feature quantity generating step, an initial state learning step, an initial condition learning step, an abnormality degree calculating step, and an unknown degree calculating step.
[0188] The state signal generation step generates a state signal ss obtained by detecting the state of the mechanical device 2 in a time series. The condition signal generation step generates a condition signal cs obtained by detecting the operating condition indicating the operating condition of the mechanical device 2 in a time series. The state feature quantity generation step generates a state feature quantity sc based on the state signal ss. The condition feature quantity generation unit 16 generates a condition feature quantity cc based on the condition signal cs. The initial state learning step outputs the result of learning based on the state feature quantity sc during initial state learning, that is, the initial learning state feature quantity lsc, as an initial state learning result slr.
[0189] The initial condition learning process outputs the result of learning based on the condition feature quantity cc during initial condition learning, i.e., the condition feature quantity lcc for initial learning, as the initial condition learning result clr. The abnormality degree calculation process obtains the initial state learning result slr or the additional state learning result aslr as the state learning result, and calculates the abnormality degree an based on the state learning result and the state feature quantity sc during detection, i.e., the state feature quantity dsc for detection. The unknown degree calculation process obtains the initial condition learning result clr or the additional condition learning result aclr as the condition learning result, and calculates the unknown degree un based on the condition learning result and the condition feature quantity cc during detection, i.e., the condition feature quantity dcc for detection. Here, it is preferred that the detection time of the condition feature quantity cc is the same as the detection time of the state feature quantity sc.
[0190] In addition, in the present embodiment, the suppression of false detection is described, but the suppression of omission can also be suppressed in the same manner as false detection. In addition, in the present invention, the case where the abnormal determination result jr is output for a normal mechanical device 2 without an abnormality is called a false detection. On the other hand, the case where the normal determination result jr is output for a mechanical device 2 with an abnormality is called an omission. In addition, in the case where the determination result is output in three stages of severe abnormality, mild abnormality, and normality, even if more than three determination results are output, the output of the erroneous determination result can be suppressed in the same manner as the suppression of false detection. As described above, according to the abnormality detection device 1 in the present embodiment, even if the operating conditions of the mechanical device 2 change, it is possible to perform abnormality detection with less output of erroneous determination results such as false detection and omission. In addition, in the case where the operating conditions of the mechanical device 2 when the initial state learning is performed are different from the operating conditions of the mechanical device 2 that is the object of the abnormality detection, the output of erroneous determination results such as false detection and omission can also be suppressed. In the method described in the present embodiment, even when the initial learning condition signal lcs and the detection condition signal dcs are different, the occurrence of the output of the wrong determination result such as false detection and omission can be suppressed. Here, in the present embodiment, the operating condition is the operating condition oc. Here, the output of the wrong determination result of the present invention includes not only the display to the operator, the output of the signal indicating the determination result, etc., but also the change of the operating state of the mechanical device 2.
[0191] According to the abnormality detection device in the present embodiment, in a system for detecting abnormalities occurring in a mechanical device 2, even for a mechanical device 2 whose operating conditions change in a complex manner, it is possible to suppress output of erroneous determination results such as false detection and omission.
[0192] Implementation method 2.
[0193] Fig.15 This is a block diagram showing an example of the structure of a mechanical system 100a involved in the present embodiment. In the mechanical system 100a, an abnormality detection device 1 of Embodiment 1 is replaced with an abnormality detection device 1a. The abnormality detection device 1a has an additional condition learning unit 22 and an additional state learning unit 23 on the basis of the structural elements of the abnormality detection device 1. In addition, in the abnormality detection device 1a, an unknown degree calculation unit 18 is replaced with an unknown degree calculation unit 18a. In addition, in the abnormality detection device 1a, an abnormality degree calculation unit 14 is replaced with an abnormality degree calculation unit 14a. In addition, in the abnormality detection device 1a, an abnormality determination unit 19a is replaced with an abnormality determination unit 19. Regarding the above-mentioned points, the abnormality detection device 1a is different from the abnormality detection device 1. Fig.15In the embodiment 1, Figure 1 The same structural elements or corresponding structural elements are denoted by the same reference numerals.
[0194] Fig.16 1 is a block diagram showing an example of the structure of the additional condition learning unit 22 involved in the present embodiment. The additional condition learning unit 22 includes a condition feature quantity storage unit 221 for storing the condition feature quantity cc, and a condition learning determination unit 222 for determining whether to perform additional condition learning, i.e., additional condition learning. In addition, the additional condition learning unit 22 includes a condition feature quantity extraction unit 223 for extracting the condition feature quantity cc, and an additional condition learning execution unit 224 for executing the additional condition learning. The additional condition learning execution unit 224 executes the additional condition learning on the detection condition feature quantity dcc extracted by the condition feature quantity extraction unit 223.
[0195] Fig.17 1 is a block diagram showing an example of the structure of the additional state learning unit 23 involved in the present embodiment. The additional state learning unit 23 includes a state feature quantity storage unit 231 for storing state feature quantities sc, and a state learning determination unit 232 for determining whether to perform additional state learning, that is, additional state learning, with respect to each unknown degree un. In addition, the additional state learning unit 23 includes a state feature quantity extraction unit 233 for extracting the state feature quantity sc, and an additional state learning execution unit 234 for executing additional state learning. The additional state learning execution unit 234 executes additional state learning with respect to the detection state feature quantity dsc extracted by the state feature quantity extraction unit 233.
[0196] exemplify Fig.16 The manner of each structural element of the additional condition learning unit 22 shown. The condition feature storage unit 221 stores the detection condition feature dcc during a certain period of time. Here, the detection condition feature dcc is output in multiple in time series. Moreover, the unknown degree calculation unit 18a outputs multiple unknown degrees un in time series based on the initial condition learning result clr and the multiple detection condition feature quantities dcc output in time series. The condition learning determination unit 222 performs a determination on whether to perform additional condition learning on each of the multiple unknown degrees un. For example, the condition learning determination unit 222 can compare the obtained unknown degree un with a predetermined threshold value (third threshold value).
[0197] The operation of the conditional learning determination unit 222 is illustrated. One of the multiple unknowns un is set as the unknown un-i (i is an integer greater than or equal to 1). Furthermore, an unknown un among the multiple unknowns un and different from the unknown un-i is set as the unknown un-j (j is an integer greater than or equal to 1 different from i). Here, i and j are respectively set as independent variables of the unknown un-i and the unknown un-j. The result of the comparison is that the unknown un-i is less than or equal to the third threshold value, and the unknown un-j is greater than the third threshold value. In this case, the conditional learning determination unit 222 outputs the independent variable i and does not output the independent variable j. The above is an illustration of the operation of the conditional learning determination unit 222.
[0198] The condition feature quantity extraction unit 223 obtains the independent variable i output by the condition learning determination unit 222. Furthermore, the detection condition feature quantity dcc-i corresponding to the obtained independent variable i is extracted from the plurality of detection condition feature quantities dcc stored in the condition feature quantity storage unit 221. The additional condition learning execution unit 224 executes condition learning based on the extracted detection condition feature quantity dcc-i. This condition learning is referred to as additional condition learning. The initial condition learning and additional condition learning described in Implementation 1 are included in the condition learning. In other words, the initial condition learning and the additional condition learning are each a mode of condition learning.
[0199] The method of additional condition learning performed by the additional condition learning execution unit 224 can be the same as the method of initial condition learning described in Implementation Example 1, except that condition learning is performed based on the detection condition feature quantity dcc instead of the initial learning condition feature quantity lcc. In addition, the modified example of initial condition learning described in Implementation Example 1 can also be applied to additional condition learning. In addition, additional condition learning can be performed in the same manner as initial condition learning, or in a different manner, but preferably in the same manner. The additional condition learning and the initial condition learning are set to the same manner, thereby calculating the unknown degree un after the additional condition learning in the same manner as the unknown degree un before the additional condition learning, so that there can be a consistency in the judgment performed by the abnormality judgment unit 19a. The aforementioned judgment performed by the abnormality judgment unit 19a is a judgment of abnormality detection based on the abnormality degree an and the unknown degree un. Here, the result of the additional condition learning is referred to as the additional condition learning result alcr. As mentioned above, the initial condition learning result clr and the additional condition learning result alcr are included in the condition learning result. In Fig.16 In the example of , the additional condition learning execution unit 224 outputs the additional condition learning result alcr-i corresponding to the independent variable i. Fig.16 The following is a description of the components of the additional condition learning unit 22.
[0200] Furthermore, the processing after the additional condition learning execution unit 224 outputs the additional condition learning result alcr will be described. Fig.15 As shown, the unknown degree calculation unit 18a updates the condition learning result from the initial condition learning result clr to the additional condition learning result alcr. Furthermore, the unknown degree calculation unit 18a calculates the unknown degree un based on the updated condition learning result, i.e., the additional condition learning result alcr, and the detection condition feature dcc obtained after the update. Fig.15 In the example of FIG. 1 , the calculation method of the unknown degree un of the unknown degree calculation unit 18a is described as being the same before and after the acquisition of the additional condition learning result alcr, except that the additional condition learning result alcr is used instead of the initial condition learning result clr. The calculation method of the unknown degree un may be changed before and after the acquisition of the additional condition learning result alcr. The operation of the abnormality determination unit 19a will be described after the description of the additional state learning unit 23.
[0201] In addition, Fig.16 In the example of , in order to associate the unknown degree un with the conditional characteristic quantity cc (in this case, the detection conditional characteristic quantity dcc), an independent variable i is set in the unknown degree un-i. However, in order to associate the unknown degree un with the conditional characteristic quantity cc, a method different from the independent variable can also be used. For example, a label, a symbol, etc. that can be attached to the data and is different from the independent variable i can be used for association. In addition, for example, the independent variable can be replaced, and a group of data corresponding to each other such as the conditional signal cs, the conditional characteristic quantity cc, and the unknown degree un can be attached with the same number and associated. In addition, the conditional characteristic quantity cc used in the calculation of the unknown degree un-i is set to the conditional characteristic quantity cc-i, and the operating condition oc used to obtain the conditional characteristic quantity cc-i is set to the operating condition oc-i. In this case, the independent variable i can be replaced and the time of obtaining the operating condition oc-i can be attached to the unknown degree un and the conditional characteristic quantity cc, thereby associating.
[0202] Next, we illustrate Fig.17 The additional state learning unit 23 is shown in FIG. The additional state learning unit 23 includes a state feature quantity storage unit 231, a state learning determination unit 232, a state feature quantity extraction unit 233, and an additional state learning execution unit 234. The state feature quantity storage unit 231 stores the detection state feature quantity dsc for a certain period of time. Fig.17In the example of , a plurality of detection state feature quantities dsc are output in time series. On the other hand, the unknown degree calculation unit 18a outputs a plurality of unknown degrees un in time series based on the initial condition learning result clr and the plurality of detection condition feature quantities dcc output in time series. The state learning determination unit 232 determines whether to perform additional state learning for each of the plurality of unknown degrees un output from the unknown degree calculation unit 18a.
[0203] The operation of the state learning determination unit 232 is illustrated. For example, the state learning determination unit 232 can compare the obtained unknown degree un with a predetermined threshold value (the fourth threshold value). One of the multiple unknown degrees un is set as the unknown degree un-m (m is an integer greater than or equal to 1). Moreover, one of the multiple unknown degrees un and different from the unknown degree un-m is set as the unknown degree un-n (n is an integer greater than or equal to 1 different from m). Here, m and n are set as the independent variables of the unknown degree un-m and the unknown degree un-n, respectively. The result of the comparison is that the unknown degree un-m is less than or equal to the fourth threshold value, and the unknown degree un-n is greater than the fourth threshold value. In this case, the conditional learning determination unit 222 outputs the independent variable m and does not output the independent variable n. The above is an example of the operation of the state learning determination unit 232.
[0204] The state feature quantity extraction unit 233 extracts a plurality of detection state feature quantities dsc stored in the state feature quantity storage unit 231 and compares the independent variable (in Fig.17 In the example of the independent variable m), the detection state feature dsc-m corresponding to the detection state feature dsc-m is extracted. The additional state learning execution unit 234 executes state learning based on the extracted detection state feature dsc-m. This state learning is called additional state learning. As mentioned above, the initial state learning described in Implementation 1 and the additional state learning described in this implementation are included in the state learning. In other words, the initial state learning and the additional state learning are each a method of state learning.
[0205] like Fig.17As shown in the example, the method of additional state learning can be the same as the method of initial state learning described in Implementation Example 1, except that the state learning is performed based on the detection state feature quantity dsc instead of the initial learning state feature quantity lsc. In addition, the additional state learning can be performed in the same manner as the initial state learning, or in a different manner, but it is preferably set to the same manner. By setting the additional state learning and the initial state learning to the same manner, the abnormality degree an after the additional state learning is calculated in the same manner as the abnormality degree an before the additional state learning, so that there can be a consistency in the judgment performed by the abnormality judgment unit 19a. The aforementioned judgment performed by the abnormality judgment unit 19a is a judgment based on abnormality detection of the abnormality degree an and the unknown degree un. In addition, the modified example of the initial state learning described in Implementation Example 1 can also be applied to the additional state learning.
[0206] Here, the result of the additional state learning is referred to as the additional state learning result aslr. The initial state learning result slr and the additional state learning result aslr are included in the state learning result. Fig.17 In the example of , the additional state learning execution unit 234 outputs the additional state learning result aslr-m corresponding to the independent variable m. Fig.17 The following is a description of each component of the additional state learning unit 23.
[0207] Furthermore, the processing related to the output additional state learning result aslr is illustrated. Fig.15 As shown, the abnormality calculation unit 14a updates the maintained state learning result from the initial state learning result slr to the additional state learning result aslr. Moreover, the abnormality calculation unit 14a outputs the abnormality an based on the updated state learning result, i.e., the additional state learning result aslr, and the detection state feature dsc obtained after the update. Here, the calculation method of the abnormality an of the abnormality calculation unit 14a may be the same or different before and after the state learning result is updated. When the calculation method of the abnormality an is set to be the same before and after the update, the abnormality an can be made consistent.
[0208] In addition, the abnormality determination unit 19a outputs the determination result jr based on the abnormality degree an and the unknown degree un, similarly to the abnormality determination unit 19 described in the first embodiment. Here, the abnormality degree an and the unknown degree un obtained by the abnormality determination unit 19a are output after the unknown degree calculation unit 18a updates the condition learning result and the abnormality degree calculation unit 14a updates the state learning result. As described above, the abnormality detection device 1a of this embodiment performs additional condition learning on the basis of initial condition learning, and performs additional state learning on the basis of initial state learning.
[0209] Here, the conditional learning determination unit 222 can further use the unknown degree un calculated using the additional conditional learning result aclr to determine whether to perform additional conditional learning, and update the conditional learning result. In addition, the state learning determination unit 232 can also further use the unknown degree un calculated using the additional state learning result aslr to determine whether to perform additional state learning, and appropriately update the state learning result. In addition, in the additional state learning unit 23, the unknown degree un and the state feature sc can be associated by means other than independent variables, which is the same as the association between the unknown degree un and the condition feature cc in the additional conditional learning unit 22.
[0210] Fig.18 is a flowchart showing an example of the operation of the additional condition learning unit 22. Fig.18 During the initial learning time before the start (START) of the initial learning, the initial condition learning unit 17 outputs the initial condition learning result clr based on the condition feature quantity lcc for initial learning. In addition, the unknown degree calculation unit 18a has the initial condition learning result clr.
[0211] The following is an explanation of the operation during the detection time. At this time, the abnormality detection device 1a is performing abnormality detection. The condition feature quantity generating unit 16 generates the detection condition feature quantity dcc based on the operating condition oc during the detection time. The detection condition feature quantity dcc is included in the condition feature quantity cc, so Fig.15 The number of the conditional feature quantity cc is shown in .
[0212] In step S2021, the condition feature storage unit 221 stores the detection condition feature dcc. In step S2022, the condition learning determination unit 222 increases the independent variable by one. For example, as time passes, the independent variable is added to the detection condition feature dcc acquired in time series. Fig.16 In step S2023, the conditional learning determination unit 222 determines whether the unknown degree un-i is greater than Fig.16 The third threshold value described in the above description may be the same value as the threshold value used by the abnormality determination unit 19a to determine whether it is appropriate to determine whether it is normal or abnormal based on the abnormality degree an, or may be a different value. The threshold value used by the abnormality determination unit 19a to determine whether it is appropriate to determine whether it is normal or abnormal based on the abnormality degree an is, for example, the threshold value used in the first embodiment. Fig.12A value such as the threshold THU1" described in the action example. When the threshold is set to the same value as the value at which the abnormality determination unit 19a makes a normal or abnormal determination based on the abnormality degree an, it is appropriate to perform additional conditional learning when the abnormality determination unit 19a is not suitable for making a normal or abnormal determination.
[0213] When the unknown degree un-i is less than or equal to the threshold value, it is determined that no additional conditional learning is required for the unknown degree un-i of the independent variable i, and the additional conditional learning unit 22 proceeds to step S2022. In this case, additional conditional learning is not performed on the independent variable i, and the conditional learning result of the unknown degree calculation unit 18a is not updated but maintained. Moreover, the calculation of the unknown degree un continues based on the conditional learning result of the unknown degree calculation unit 18a and the conditional feature quantity dcc for detection. In this case, the addition of the independent variable is performed again in step S2022, and the conditional learning determination unit 222 determines whether to implement additional conditional learning for the unknown degree un of the updated independent variable i+1.
[0214] In step S2023, when the unknown degree un-i is greater than the threshold, the process proceeds to step S2024. In step S2024, the conditional feature quantity extraction unit 223 obtains the independent variable i, and extracts the conditional feature quantity cc corresponding to the independent variable i, in other words, the detection conditional feature quantity dcc-i, from the conditional feature quantity storage unit 221. Furthermore, the additional condition learning unit 22 proceeds to step S2025. In step S2025, the additional conditional learning execution unit 224 outputs the additional conditional learning result aclr-i based on the detection conditional feature quantity dcc-i extracted by the conditional feature quantity extraction unit 223.
[0215] When the process proceeds to step S2025, the unknown degree calculation unit 18a updates the conditional learning result held so far to the additional conditional learning result aclr-i. Fig.18 1 and 2. The operation flow of the additional condition learning unit 22 is shown in FIG. 1. After the condition learning result is updated, the unknown degree calculation unit 18a calculates the unknown degree un based on the updated condition learning result and the detection condition feature value dcc obtained after the update. The processing of the unknown degree calculation unit 18a is performed on the detection condition feature value dcc generated in the condition feature value generation unit 16 until the condition learning result is updated next.
[0216] In addition, Fig.18In the example shown, the additional conditional learning execution unit 224 outputs the additional conditional learning result aclr-i based on the detection conditional feature dcc-i. However, the present embodiment is not limited to the above-described method. The number of detection conditional feature dcc used in the additional conditional learning and the independent variable of the detection conditional feature dcc can be freely selected. For example, multiple data obtained after the detection conditional feature dcc-i can be selected. If an example is given, 100 data from the detection conditional feature dcc-i+1 to the detection conditional feature dcc-i+100 are extracted. Moreover, the additional conditional learning result aclr-i can be output based on the extracted 100 data. Moreover, during the period of extracting data, the conditional learning determination unit 222 can be configured to determine that the conditional learning result is not to be updated.
[0217] In addition, for example Fig.18 The condition feature quantity storage unit 221 stores the initial learning condition feature quantity lcc used for initial condition learning. Fig.18 In step S2025, the additional condition learning execution unit 224 may execute additional condition learning based on the stored initial learning condition feature quantity lcc and the detection condition feature quantity dcc-i acquired thereafter. The additional condition learning execution unit 224 uses the initial learning condition feature quantity lcc as part of the learning data for additional condition learning, thereby making up for the lack of data and performing learning even when the amount of the acquired detection condition feature quantity dcc is insufficient.
[0218] In addition, the condition learning determination unit 222 may be replaced by the abnormality determination unit 19a to determine whether to perform additional condition learning. In other words, when the abnormality determination unit 19a determines that the unknown degree un is greater than the threshold value, the additional condition learning unit 22 may perform additional condition learning. In the above-mentioned manner, when the abnormality determination unit 19a is in an unknown state that is not suitable for determining whether it is normal or abnormal, additional condition learning is performed, so that abnormality detection can be performed efficiently. In addition, in the case of the above-mentioned manner, the condition learning determination unit 222 may be omitted.
[0219] In addition, the conditional learning determination unit 222 only needs to determine whether to perform additional conditional learning based on the unknown degree un to be performed, and the method is not limited to using Fig.18The method described. For example, a threshold is set for the unknown degree un. Moreover, it is possible to perform a judgment on whether the unknown degree un exceeds the threshold for each of the multiple un obtained in a time series, and when the unknown degree un exceeds the threshold continuously for a predetermined number of times, it is determined that additional conditional learning is to be performed. In addition, the additional conditional learning execution unit 224 can obtain multiple detection conditional feature quantities dcc corresponding to the unknown degree un exceeding the threshold, and implement additional conditional learning based on the obtained multiple detection conditional feature quantities dcc. In the manner described above, if a judgment is made on whether to perform additional conditional learning, it is possible to judge whether additional conditional learning needs to be performed based on the values of multiple unknown degrees un continuously in a time series, so there is an advantage that it is not easy to make an erroneous judgment on whether additional conditional learning needs to be performed.
[0220] As described above, by implementing additional condition learning, when unknown operating conditions OC, detection condition characteristic quantities DCC, etc. are calculated, the abnormality detection device 1a obtains information of unknown operating conditions OC, detection condition characteristic quantities DCC, etc. Moreover, it is possible to perform abnormality detection corresponding to the unknown operating conditions OC, detection condition characteristic quantities DCC, etc. Thus, it is possible to perform abnormality detection with less output of erroneous determination results such as false detection and omission for various operating conditions OC and various detection condition characteristic quantities DCC.
[0221] Fig.19 is a flowchart showing an example of the operation of the additional state learning unit 23 according to the present embodiment. Fig.19 During the initial learning time before the start (START) of the initial state learning unit 13, the initial state learning unit 13 outputs the initial state learning result slr based on the initial learning state feature quantity lsc. In addition, the abnormality degree calculation unit 14a has the initial state learning result slr.
[0222] The following describes the operation of the detection time after the initial learning time. At this time, the abnormality detection device 1a is performing abnormality detection. The state feature quantity generating unit 12 generates the detection state feature quantity dsc based on the state quantity sa during the detection time. The detection state feature quantity dsc is included in the state feature quantity sc, so Fig.15 In FIG. 1 , the number of the state feature quantity sc is shown.
[0223] In step S2151, the state feature storage unit 231 stores the detection state feature dsc. In step S2152, for example, the state learning determination unit 232 increases the independent variable by one. For example, as time passes, the independent variable is sequentially added to the detection state feature dsc acquired in time series. Fig.17The independent variable m-1 described in the above is updated to the independent variable m. In step S2153, the state learning determination unit 232 determines whether the unknown degree un-m is greater than a predetermined threshold value. The threshold value may be the same as the threshold value used by the abnormality determination unit 19a when determining whether it is abnormal or normalized based on the abnormality degree an, or it may be a different value. In the case where the threshold value of the state learning determination unit 232 is set to the same value as the threshold value used when determining whether it is abnormal or normalized based on the abnormality degree an, in the case where the abnormality determination unit 19a is not suitable for normal or abnormal determination, it is suitable because additional state learning is performed. In addition, the threshold value of the state learning determination unit 232 may be set to the same value as the threshold value of the conditional learning determination unit 222, or it may be a different value. In the case of setting it to the same value, one of the conditional learning determination unit 222 and the state learning determination unit 232 can be omitted. In addition, the calculation load can be reduced. Moreover, in the case of performing additional conditional learning, additional state learning is also performed, so the state learning result and the conditional learning result are updated at the same time. In the case as described above, the abnormality degree calculation unit 14 a uses the updated state learning result that matches the updated condition learning result used by the unknown degree calculation unit 18 a , and thus can perform abnormality detection with higher accuracy.
[0224] When the unknown degree un-m is less than or equal to a predetermined threshold value, it is determined that no additional state learning is required for the unknown degree un-m of the independent variable m, and the additional state learning unit 23 proceeds to step S2152. In this case, no additional state learning is performed for the independent variable m, and the state learning result used by the abnormality calculation unit 14a is not updated but maintained. Moreover, the calculation of the abnormality an continues based on the state learning result maintained by the abnormality calculation unit 14a and the detection state feature dsc. Again, in step S2152, the independent variable is updated from m to m+1, and the state learning determination unit 232 determines whether to implement additional state learning for the next independent variable m+1.
[0225] In step S2153, when the unknown degree un-m is greater than the threshold, the additional state learning unit 23 proceeds to step S2154. In step S2154, the state feature extraction unit 233 obtains the independent variable m, and extracts the state feature sc corresponding to the independent variable m, in other words, the detection state feature dsc-m, from the state feature storage unit 231. Moreover, the additional state learning unit 23 proceeds to step S2155. In step S2155, the additional state learning execution unit 234 outputs the additional state learning result aslr-m based on the detection state feature dsc-m extracted by the state feature extraction unit 233. The above is Fig.19 The operation flow of the additional state learning unit 23 is shown.
[0226] When entering step S2155, the abnormality calculation unit 14a updates the state learning result maintained so far to the additional state learning result aslr-m. After the update of the state learning result is executed, the abnormality calculation unit 14a calculates the abnormality an based on the updated state learning result and the detection state feature dsc obtained after the update. The processing of the abnormality calculation unit 14a that calculates the abnormality an based on the updated state learning result and the detection state feature dsc obtained after the update is executed respectively for the detection state feature dsc generated in the state feature generation unit 12 until the state learning result is updated next.
[0227] In addition, Fig.19 In the example shown, the additional state learning execution unit 234 outputs the additional state learning result aslr-m based on the detection state feature dsc-m, but it is not limited to the above-mentioned method. The state feature sc used for the additional state learning can be freely selected. For example, 100 detection condition feature values dcc generated after the detection state feature value dsc-m can be used for additional state learning in a manner from the detection state feature value dsc-m+1 to the detection state feature value dsc-m+100. The additional state learning result aslr-m can be output based on them.
[0228] When the additional state learning unit 23 is configured to perform additional state learning on the unknown degree un obtained by performing additional condition learning in the additional condition learning unit 22, it is possible to update not only the condition learning result maintained by the unknown degree calculation unit 18a, but also the state learning result maintained by the abnormality calculation unit 14a. Therefore, more accurate abnormality detection can be achieved. It is possible to achieve abnormality detection with fewer outputs of erroneous judgment results such as false detection and omission. In addition, it can also be configured to only perform additional condition learning in the additional condition learning unit 22, and not update the state learning result in the abnormality calculation unit 14a. In other words, the structure in which the additional state learning unit 23 is omitted and the abnormality calculation unit 14a calculates the abnormality an based on the initial state learning result slr and the detection state feature dsc can be maintained. In the structure as described above, the condition learning result is updated when the operating condition OC is determined to be unknown. Therefore, compared with a structure that does not have an additional condition learning unit 22, such as the abnormality detection device 1 described in embodiment 1, it is possible to perform abnormality detection with less output of erroneous judgment results such as false detection and omission.
[0229] Next, in order to explain the effect of additional conditional learning, a comparison is made between a configuration in which the additional state learning unit 23 is omitted from the abnormality detection device 1a and a configuration in which the additional state learning unit 23 and the additional conditional learning unit 22 are omitted from the abnormality detection device 1a. In the following description, as an example of a configuration in which the additional state learning unit 23 and the additional conditional learning unit 22 are omitted from the abnormality detection device 1a, a configuration in which the additional state learning unit 23 and the additional conditional learning unit 22 are omitted is described. Figure 1 The structure and operation of the abnormality detection device 1 according to the first embodiment will be described.
[0230] Fig. 20 1 is a diagram showing an example of time changes of the abnormality degree an, the unknown degree un, and the determination result jr generated by the configuration in which the additional condition learning unit 22 and the additional state learning unit 23 are omitted from the abnormality detection device 1a according to the present embodiment. Fig. 20 In the embodiment, the detection result of the abnormality detection device 1 described in the first embodiment is shown as an example of a structure without the additional condition learning unit 22 and the additional state learning unit 23. Fig.21 This is a diagram showing an example of time changes of the abnormality degree an, the unknown degree un, and the determination result jr generated by the configuration in which the additional state learning unit 23 is omitted from the abnormality detection device 1 a according to the present embodiment. Fig.21 This is the result obtained by detecting the state of the mechanical device 2 by omitting the additional state learning unit 23 from the abnormality detection device 1a. Fig. 20 and Fig.21 comparison.
[0231] Get Fig. 20 The state of the mechanical device 2 at the time of the data shown is assumed to be that the mechanical device 2 does not deteriorate until the time td2, and the mechanical device 2 gradually deteriorates after the time td2. In addition, between the time tb2 and the time tc2, and between the time td2 and the time tg2, the speed of the motor is changed to a setting different from that at the time of initial learning. Moreover, during the time from the time ta2 to the time tb2, and the time from the time tc2 to the time td2, the operating condition oc, that is, the speed of the motor is the same as that at the time of initial learning.
[0232] Fig. 20 (a) and Fig.21 (a) shows the temporal change of the abnormality an. The data points of the abnormality an are plotted in units of one hour. Fig. 20 (b) and Fig.21 (b) shows the time variation of the unknown degree un. Fig. 20 (c) and Fig.21 (c) shows the temporal change of the determination result jr. Fig. 20 and Fig.21The horizontal axis represents time, in hours. Fig. 20 (a) to Fig. 20 In the time axis of (c), positions with the same number represent the same time. Fig.21 (a) to Fig.21 In the time axis of (c), positions with the same number represent the same time.
[0233] exist Fig. 20 In the example of , data from the time ta2 to the time tg2, which is the time when the time ta2 has passed since the start of the operation of each abnormality detection device, is shown. Fig. 20 In the illustrated example, initial condition learning and initial state learning are performed between the start time of the operation of each abnormality detection device and the time ta2 at which the detection time starts.
[0234] exist Fig. 20 In (a), although the abnormality an from time ta2 to time td2 contains a certain amount of fluctuation, it is all plotted between 0 and 1 and generally changes little. Fig. 20 In (a), from time td2 to time tg2, the abnormality an gradually increases as the operating time passes. Fig. 20 In (a), the threshold value THF2 related to the abnormality an is set to 1. The threshold value THF2 is determined based on the distribution of the abnormality an when the mechanical device 2 is normal. Fig. 20 In (a), for ease of understanding, the true abnormality TRUE2 of the mechanical device 2 is indicated by a solid line. The description of the true abnormality TRUE2 is the same as in Fig.10 The description of the true abnormality TRUE1 described in is the same as that described in , so it is omitted.
[0235] exist Fig. 20 In (b), the unknown degree un is plotted between 0 and 1 from time ta2 to time tb2 and from time tc2 to time td2. On the other hand, the unknown degree un is plotted between 1 and 2 from time tb2 to time tc2 and from time td2 to time tg2. Fig. 20 In (b), the threshold value THU2 related to the unknown degree un is set to 1. In addition, Fig. 20 The judgment result jr of (c) is plotted by connecting the data points in units of one hour with a line. When the mechanical device 2 is normal, the judgment result jr shows a value of 0. In addition, when the mechanical device 2 is abnormal, the judgment result jr shows a value of 1. In addition, the output method of the judgment result jr is not limited to the above method.
[0236] exist Fig. 20 In the example of Fig.12 Similarly to the description of , the judgment result jr is output. Therefore, if Fig. 20 As shown in (b), the determination result jr has an abnormality an of 1 or less in the entire period from time ta2 to time tg2, so the determination result jr becomes a value of 0 indicating normality. Fig. 20 At the time after time td2 in (a), the true abnormality degree TRUE2 gradually increases.
[0237] As mentioned above, according to Fig. 20 The determination result jr of the abnormality detection device 1, i.e., the determination result jr of the abnormality detection device 1, has omitted the abnormal state of the mechanical device 2 from being determined as a normal state. This omission occurs because, after time td2, an operating condition oc different from the operating condition oc during initial learning is applied to the mechanical device 2. In other words, the omission occurs because the detection condition signal dcs and the initial learning condition signal lcs are different.
[0238] Next, Fig.21 The following example is used to illustrate this. Fig.21 In the example of , data from the time ta2' to the time tg2', which is the time when the abnormality detection device starts to operate, is shown. Fig.21 In the example shown, initial condition learning and initial state learning are performed between the start time of the operation of each abnormality detection device and the time ta2' at which the detection time starts. Fig.21 In the example, after time td2', the mechanical device 2 gradually deteriorates. Fig.21 In the time from time tb2' to time tc2' and the time from time td2' to time tg2', the speed of the motor is changed to a setting different from that in the initial learning. That is, the operating condition oc during the detection is changed relative to the operating condition oc during the initial learning. Moreover, the speed of the motor is set to the same as that in the initial learning during the time from time ta2' to time tb2' and the time from time tc2' to time td2'. That is, the operating condition oc is set to be the same as the operating condition oc during the initial learning.
[0239] exist Fig.21 In (a), time ta2' is the time when time ta2' has passed after the abnormality detection device 1a starts to operate. Also, time tg2' is the time when time tg2' has passed after the abnormality detection device 1a starts to operate. In addition, it is assumed that the initial state learning and initial condition learning of the mechanical device 2 in the normal state are completed before time ta2'. Fig.21In (a), data points of the abnormality degree an are plotted in units of one hour.
[0240] exist Fig.21 In (a), the value of the abnormality an is plotted between 0 and 1 between the time ta2' and the time tb2', and between the time tb2' and the time td2'. Fig.21 In (a), at the time point tb2', the abnormality an is plotted as a value Fv. The value Fv becomes a value exceeding 1. Fig.21 In (a), from time td2' to time tg2', the abnormality degree an increases as the operating time passes.
[0241] exist Fig.21 In (a), the threshold value THF2' of the abnormality an is set to 1. The threshold value THF2' is a predetermined threshold value. The threshold value THF2' can be determined based on the distribution of the abnormality an when the mechanical device 2 is in a normal state. Fig.21 In (a), based on the abnormality an estimated by the abnormality detection device 1a, the true abnormality TRUE2' of the mechanical device 2 which is not affected by the disturbance of the operating condition oc is indicated by a bold line. The description of the true abnormality TRUE2' is the same as that in the first embodiment. Fig.10 The description about the true abnormality TRUE1 described in is the same as that described in , so it is omitted.
[0242] In addition, Fig.21 In (b), the unknown degree un is plotted between 0 and 1 during the period from time ta2' to time tb2' and from time tb2' to time td2'. On the other hand, at the time point of time tb2', the unknown degree un is plotted as a position value Uv. Position Uv shows that at time tb2', the value of the unknown degree un becomes a value exceeding 1.
[0243] exist Fig.21 In the example of , at the time point of time tb2', the additional condition learning unit 22 performs additional condition learning. Therefore, in the time after time tb2', the increase of the unknown degree un is suppressed. On the other hand, Fig. 20 In the example of , the condition learning unit 22 is not added, and the condition learning result is not updated. Therefore, the unknown degree un based on the initial condition learning result clr continues to be output in the time after time tb2. Therefore, in the time from time tb2 to tc2 and from time td2 to time tg2, the unknown degree un becomes a value greater than 1.
[0244] like Fig.21As shown in (b), in the time after time tb2', the increase of the unknown degree un is suppressed, and as a result, Fig.21 (c) The judgment result jr and Fig. 20 (c) Unlike the above, an increase in the abnormality degree an of the mechanical device 2 can be accurately detected as an abnormality.
[0245] As described above, the abnormality detection device 1a determines whether additional condition learning is required based on the unknown degree un. Moreover, the condition learning result can be updated when necessary. Therefore, even when the operating condition oc changes, it is possible to perform abnormality detection with less false detection and omission by updating the condition learning result. In addition, the abnormality detection device 1a determines whether additional state learning is required based on the unknown degree un. Moreover, it is possible to update the state learning result when necessary. Therefore, even when the state quantity sa changes with the change of the operating condition oc, it is possible to perform abnormality detection with less false detection and omission by updating the state learning result. Through these structures, the abnormality detection device 1a can suppress false detection and omission even when the operating condition oc of the mechanical device 2 changes. In addition, the abnormality detection device 1a determines whether to perform additional condition learning, additional state learning, etc. based on the detection condition signal dcs of the time series. As a result, even when the operating condition oc changes in a complex manner depending on time, it is possible to accurately determine whether additional condition learning, additional state learning, etc. are required. In addition, when additional condition learning and additional state learning are performed using the time series detection condition signal dcs, the abnormality detection device 1a can also calculate accurate unknown degree un, abnormality degree an, etc. when the operating condition oc changes in a complex manner depending on time. Therefore, the output of erroneous judgment results such as false detection and omission can be suppressed, and the abnormality can be detected with high accuracy. In addition, in this embodiment, the suppression of false detection and omission is described, but it is not limited to this. For example, similar to embodiment 1, when the judgment result is output in three stages of severe abnormality, mild abnormality, and normal, even when the output is greater than or equal to three judgment results, the output of erroneous judgment results can be suppressed.
[0246] The abnormality detection device 1a of this embodiment may further include an additional condition learning unit 22 in addition to the components of the abnormality detection device 1 described in Embodiment 1. The additional condition learning unit 22 includes a condition feature quantity storage unit 221, a condition learning determination unit 222, a condition feature quantity extraction unit 223, and an additional condition learning execution unit 224.
[0247] The condition feature quantity storage unit 221 stores the detection condition feature quantity dcc. The condition learning determination unit 222 determines whether to perform additional condition learning based on the unknown degree un. The condition feature quantity extraction unit 223 extracts the detection condition feature quantity dcc used for additional condition learning from the condition feature quantity storage unit 221 when the condition learning determination unit 222 determines to perform additional condition learning. The additional condition learning execution unit 224 outputs the result of executing additional condition learning based on the extracted detection condition feature quantity dcc as an additional condition learning result aclr.
[0248] The abnormality detection method of this embodiment may further include an additional condition learning step in addition to the steps of the abnormality detection method described in Embodiment 1. The additional condition learning step includes a condition feature quantity storage step, a condition learning determination step, a condition feature quantity extraction step, and an additional condition learning execution step.
[0249] The condition feature quantity storage step stores the detection condition feature quantity dcc. The condition learning determination step determines whether to perform additional condition learning based on the unknown degree un. The condition feature quantity extraction step extracts the detection condition feature quantity dcc used for additional condition learning from the detection condition feature quantity dcc stored in the condition feature quantity storage step when it is determined in the condition learning determination step that additional condition learning is to be performed. The additional condition learning execution step outputs the result of executing additional condition learning based on the extracted detection condition feature quantity dcc as an additional condition learning result aclr.
[0250] The unknown degree calculation unit 18a may update the conditional learning result that holds the conditional learning result to the additional conditional learning result aclr when the additional conditional learning result aclr is outputted by the additional conditional learning execution unit 224. Furthermore, after executing the update of the conditional learning result, the unknown degree calculation unit 18a may calculate the unknown degree un based on the updated conditional learning result and the detection conditional feature quantity dcc outputted from the conditional feature quantity generation unit after the update.
[0251] The condition learning determination unit 222 determines to execute additional condition learning when the unknown degree un exceeds a third threshold value determined in advance, and determines not to execute additional condition learning when the unknown degree un is less than or equal to the third threshold value determined in advance.
[0252] The conditional learning determination unit 222 may determine to execute additional conditional learning only when a plurality of unknown degrees un exceed a predetermined threshold value a predetermined number of times consecutively in time series.
[0253] The abnormality detection device 1a of this embodiment may further include an additional state learning unit 23 in addition to the components of the abnormality detection device 1 described in Embodiment 1. The additional state learning unit 23 includes a state feature quantity storage unit 231, a state learning determination unit 232, a state feature quantity extraction unit 233, and an additional state learning execution unit 234.
[0254] The state feature storage unit 231 stores the detection state feature dsc. The state learning determination unit 232 determines whether to perform additional state learning based on the unknown degree un. When the state learning determination unit 232 determines to perform additional state learning, the state feature extraction unit 233 extracts the detection state feature dsc used for additional state learning from the state feature storage unit 231. The additional state learning execution unit 234 outputs the result of executing additional state learning based on the extracted detection state feature dsc as an additional state learning result aslr.
[0255] The abnormality detection method of this embodiment may further include an additional state learning step in addition to the steps of the abnormality detection method described in Embodiment 1. The additional state learning step includes a state feature quantity storage step, a state learning determination step, a state feature quantity extraction step, and an additional state learning execution step.
[0256] The state feature quantity storage step stores the detection state feature quantity dsc. The state learning determination step determines whether to perform additional state learning based on the unknown degree un. The state feature quantity extraction step extracts the detection state feature quantity dsc used for additional state learning from the detection state feature quantity dsc stored in the state feature quantity storage step when it is determined in the state learning determination step that additional state learning is to be performed. The additional state learning execution step outputs the result of performing additional state learning based on the extracted detection state feature quantity dsc as an additional state learning result aslr.
[0257] As described above, according to the present embodiment, it is possible to provide an abnormality detection device with less false detection and omission when detecting the state of the mechanical device 2 whose operating conditions change. In addition, additional condition learning can be performed when necessary, thereby updating the condition learning result maintained by the unknown degree calculation unit 18a. In addition, additional state learning can be performed when necessary, thereby updating the condition learning result maintained by the abnormality calculation unit 14a. As a result, when detecting the state of the mechanical device 2 whose operating conditions change, more accurate abnormality detection can be performed. In addition, the occurrence of false detection and omission can be reduced.
[0258] Description of the label
[0259] 1. 1a abnormality detection device, 2 mechanical device, 3 control device, 11 state signal generating unit, 12 state feature quantity generating unit, 13 initial state learning unit, 14 abnormality degree calculating unit, 15 condition signal generating unit, 16 condition feature quantity generating unit, 17 initial condition learning unit, 18 unknown degree calculating unit, 19 abnormality determination unit, 20 motor, 22 additional condition learning unit, 23 additional state learning unit, 201 ball screw, 202 movable part, 203 guide part, 204 ball screw shaft, 202 coupling, 203 servo motor shaft, 204 servo motor, 205 encoder, 221 condition feature quantity storage unit, 2 22 conditional learning determination unit, 223 conditional feature quantity extraction unit, 224 additional conditional learning execution unit, 231 state feature quantity storage unit, 232 state learning determination unit, 233 state feature quantity extraction unit, 234 additional state learning execution unit, 310 current sensor, 311 driver, 301 PLC, 401 PC, 402 display for PLC, 403 display for PC, cc conditional feature quantity, clr conditional learning result, cs conditional signal, df driving force, jr determination result, oc operating condition, electric power pw, state quantity sa, state feature quantity sc, state learning result slr, state signal ss.
Claims
1. An abnormality detection device, comprising: a state signal generating unit for generating a state signal obtained by detecting the state of the mechanical device in a time series; a condition signal generating unit for generating a condition signal by detecting an operating condition indicating an operating condition of the mechanical device in a time series; a state feature quantity generating unit that generates a state feature quantity based on the state signal; a condition feature quantity generating unit that generates a condition feature quantity based on the condition signal; an initial state learning unit that outputs a result obtained by learning based on the state feature quantity during initial state learning, that is, the state feature quantity for initial learning, as an initial state learning result; an initial condition learning unit that outputs a result obtained by learning based on the condition feature quantity during initial condition learning, that is, the initial learning condition feature quantity as an initial condition learning result; an abnormality degree calculation unit that obtains the initial state learning result or the additional state learning result as a state learning result, and calculates the abnormality degree based on the state learning result and the state feature quantity at the time of detection, that is, the detection state feature quantity; as well as The unknown degree calculation unit obtains the initial condition learning result or the additional condition learning result as the condition learning result, and calculates the unknown degree based on the condition learning result and the condition feature quantity during the detection, that is, the detection condition feature quantity.
2. The abnormality detection device according to claim 1, wherein: An abnormality determination unit is provided, the abnormality determination unit detecting abnormality of the mechanical device based on the abnormality degree and the unknown degree.
3. The abnormality detection device according to claim 2, wherein: The abnormality determination unit determines that the state of the mechanical device is abnormal when the abnormality degree is greater than a predetermined first threshold value and the unknown degree is less than a predetermined second threshold value.
4. The abnormality detection device according to any one of claims 1 to 3, characterized in that: The condition feature quantity generation unit generates, as the condition feature quantity, a plurality of statistics calculated from the condition signal at a plurality of time points.
5. The abnormality detection device according to any one of claims 1 to 4, characterized in that: The conditional feature quantity generating unit generates, as the conditional feature quantity, a frequency characteristic obtained by frequency analysis of the time-series conditional signal.
6. The abnormality detection device according to any one of claims 1 to 5, characterized in that: The mechanical device is driven by an electric motor to perform operations. The operating condition is a control signal that defines a shape of a time response of at least any one of the position of the motor, the speed of the motor, the acceleration of the motor, the jerk of the motor, and the driving force of the motor.
7. The abnormality detection device according to any one of claims 1 to 6, characterized in that: With additional conditional learning department, The additional condition learning unit comprises: a condition feature quantity storage unit for storing the detection condition feature quantity; a condition learning determination unit that determines whether to perform additional condition learning based on the unknown degree; a condition feature quantity extraction unit that extracts the detection condition feature quantity used for the additional condition learning from the condition feature quantity storage unit when the condition learning determination unit determines that the additional condition learning is to be performed; as well as The additional condition learning execution unit outputs a result obtained by executing the additional condition learning based on the extracted detection condition feature quantity as the additional condition learning result.
8. The abnormality detection device according to claim 7, characterized in that: When the additional condition learning execution unit outputs the additional condition learning result, the unknown degree calculation unit updates the condition learning result to the additional condition learning result output from the maintained condition learning result, and after executing the update of the condition learning result, calculates the unknown degree based on the updated condition learning result and the detection condition feature output from the condition feature generation unit after the update.
9. The abnormality detection device according to claim 7, characterized in that: The conditional learning determination unit determines to execute the additional conditional learning when the unknown degree exceeds a third threshold value determined in advance, and determines not to execute the additional conditional learning when the unknown degree is less than or equal to the third threshold value determined in advance.
10. The abnormality detection device according to claim 7 or 8, characterized in that: The condition learning determination unit determines to execute the additional condition learning only when the unknown degree exceeds a predetermined threshold value a predetermined number of times continuously in a time series.
11. The abnormality detection device according to claim 7, characterized in that: With additional state learning unit, The additional state learning unit has: a state feature quantity storage unit for storing the detection state feature quantity; a state learning determination unit that determines whether to perform additional state learning based on the unknown degree; a state feature quantity extraction unit that extracts the detection state feature quantity used for the additional state learning from the state feature quantity storage unit when the state learning determination unit determines that the additional state learning is to be performed; as well as The additional state learning execution unit outputs a result obtained by executing the additional state learning based on the extracted detection state feature amount as the additional state learning result.
12. A mechanical system comprising: Mechanical devices; a state signal generating unit for generating a state signal obtained by detecting the state of the mechanical device in a time series; a condition signal generating unit for generating a condition signal by detecting an operating condition indicating an operating condition of the mechanical device in a time series; a state feature quantity generating unit that generates a state feature quantity based on the state signal; a condition feature quantity generating unit that generates a condition feature quantity based on the condition signal; an initial state learning unit that outputs a result obtained by learning based on the state feature quantity during initial state learning, that is, the state feature quantity for initial learning, as an initial state learning result; an initial condition learning unit that outputs a result obtained by learning based on the condition feature quantity during initial condition learning, that is, the initial learning condition feature quantity as an initial condition learning result; an abnormality degree calculation unit that obtains the initial state learning result or the additional state learning result as the state learning result, and calculates the abnormality degree based on the state learning result and the state feature quantity at the time of detection, that is, the detection state feature quantity; as well as The unknown degree calculation unit obtains the initial condition learning result or the additional condition learning result as the condition learning result, and calculates the unknown degree based on the condition learning result and the condition feature quantity at the time of the detection, that is, the detection condition feature quantity.
13. A method for detecting anomalies, comprising: a state signal generating step of generating a state signal obtained by detecting the state of the mechanical device in a time series; a condition signal generating step of generating a condition signal by detecting an operating condition indicating an operating condition of the mechanical device in a time series; A state feature quantity generating step of generating a state feature quantity based on the state signal; A condition feature quantity generating step of generating a condition feature quantity based on the condition signal; An initial state learning step of outputting a result obtained by learning based on the state feature quantity during initial state learning, i.e., the state feature quantity for initial learning, as an initial state learning result; An initial condition learning step of outputting a result obtained by learning based on the condition feature quantity during initial condition learning, i.e., the condition feature quantity for initial learning, as an initial condition learning result; an abnormality degree calculation step of obtaining the initial state learning result or the additional state learning result as the state learning result, and calculating the abnormality degree based on the state learning result and the state feature quantity at the time of detection, i.e., the detection state feature quantity; as well as The unknown degree calculation step obtains the initial condition learning result or the additional condition learning result as the condition learning result, and calculates the unknown degree based on the condition learning result and the condition feature quantity during the detection, that is, the detection condition feature quantity.
