Robot failure sign detection device and robot failure sign detection method

Abstract

In this robot malfunction sign detection device, a behavioral time series data acquisition unit performs a process for acquiring behavioral time series data about a motor of a joint of the robot, from a robot motion, every data collection unit period. An evaluation value calculation unit calculates an evaluation value for the behavioral time series data. A representative evaluation value generation unit generates a representative evaluation value representing the evaluation value every data collection unit period. A sequence processing unit generates a sequence composed of representative evaluation values. A determination unit creates a determination model on the basis of an initial sequence, at an initial stage of operation of the robot. The determination unit, after the initial stage of operation, inputs determination data including data based on a robot motion after the initial stage of operation into the determination model to acquire the specificity of the determination data.

Description

[Technical Field]

[0001] This application relates to monitoring the state of a robot. [Background technology]

[0002] When industrial robots are repeatedly operated in factories, it is inevitable that the various parts of the robot (for example, mechanical components) will deteriorate. As this situation progresses, the robot will eventually break down. If a robot breaks down and the line is stopped for an extended period of time, this will result in significant losses, so there is a strong demand for carrying out maintenance (preservation) before a robot breaks down. However, carrying out frequent maintenance is difficult due to factors such as maintenance costs.

[0003] In order to perform maintenance at an appropriate time, a device has been proposed for predicting the remaining life of a robot's reducer, etc. Patent Document 1 discloses this type of robot maintenance support device.

[0004] The robot maintenance support device of Patent Document 1 is configured to diagnose the future change trend of the current command value based on data on the current command value of the servo motor that constitutes the robot drive system, and determine the period until the current command value reaches a preset value based on the diagnosed change trend. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2016-117148 A Summary of the Invention [Problem to be solved by the invention]

[0006] The configuration of Patent Document 1 exemplified the I monitor, duty, and peak current as diagnostic items for the current command value. Simply using these items was not necessarily effective in detecting signs of robot failure. Therefore, a new configuration was needed that could accurately detect when a robot was approaching a failure.

[0007] The present application has been made in view of the above circumstances, and its purpose is to accurately detect signs of robot failure. [Means for solving the problem]

[0008] The problem to be solved by the present application is as described above. Next, the means for solving this problem and the effects thereof will be explained.

[0009] According to a first aspect of the present application, there is provided a robot failure sign detection device having the following configuration. Specifically, the robot failure sign detection device includes a behavior time-series data acquisition unit, an evaluation value calculation unit, a representative evaluation value generation unit, a series processing unit, and a judgment unit. The behavior time-series data acquisition unit performs a process of acquiring behavior time-series data related to motors of the robot's joints from robot motion for each data collection unit period. The evaluation value calculation unit calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit. The representative evaluation value generation unit generates a representative evaluation value representative of the evaluation values ​​from the evaluation values ​​obtained by the evaluation value calculation unit for each data collection unit period. The series processing unit generates a series of the representative evaluation values. The judgment unit creates a judgment model based on an initial series, which is a series generated by the series processing unit, at an early stage of robot operation. After the early stage of operation, the judgment unit inputs judgment data including data based on robot motion after the early stage of operation to the created judgment model and acquires the specificity of the judgment data.

[0010] According to a second aspect of the present application, there is provided a robot failure sign detection method as follows. Specifically, this robot failure sign detection method includes a behavior time-series data acquisition step, an evaluation value calculation step, a representative evaluation value generation step, a series processing step, a model creation step, and a judgment step. In the behavior time-series data acquisition step, a process of acquiring behavior time-series data related to motors of the robot's joints from robot operation is performed for each data collection unit period. In the evaluation value calculation step, an evaluation value is calculated for the behavior time-series data acquired in the behavior time-series data acquisition step. In the representative evaluation value generation step, a representative evaluation value representative of the evaluation values ​​obtained in the evaluation value calculation step is generated for each data collection unit period. In the series processing step, a series of the representative evaluation values ​​is generated. In the model creation step, a judgment model is created based on an initial series, which is a series generated in the series processing step, at the beginning of operation of the robot. In the judgment process, after the initial stage of operation, judgment data including data based on robot operation after the initial stage of operation is input into the created judgment model to obtain the specificity of the judgment data.

[0011] This makes it possible to easily grasp the signs of a robot failure, and therefore to carry out maintenance on the robot before a failure occurs. [Effects of the Invention]

[0012] According to the present application, it is possible to effectively detect signs of robot failure. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a perspective view showing a configuration of a robot according to the present application. [Figure 2] 1 is a block diagram illustrating an electrical configuration of a robot failure sign detection device according to a first embodiment. [Figure 3] 6 is a graph illustrating the timing of a trigger signal for acquiring time series data. [Figure 4] A conceptual diagram showing a hidden Markov model. [Figure 5] A conceptual diagram showing an Ergodic Hidden Markov Model with two states. [Figure 6] FIG. 2 is a schematic diagram illustrating an initial sequence and a judgment sequence. [Figure 7] 10 is a graph showing the root mean square of current values. [Figure 8] 10 is a graph showing the logarithmic likelihood of the similarity of the root mean square series of current values. [Figure 9] FIG. 10 is a block diagram illustrating the electrical configuration of a robot failure sign detection device according to a second embodiment. [Figure 10] FIG. 11 is a conceptual diagram showing the DTW method used in the third embodiment. [Figure 11] A graph showing the log-likelihood evaluated by inputting data processed by DTW into a hidden Markov model. [Figure 12] FIG. 10 is a block diagram illustrating the electrical configuration of a robot failure sign detection device according to a fourth embodiment. [Figure 13] FIG. 13 is a schematic diagram illustrating an initial sequence and determination data in the fourth embodiment. [Figure 14] FIG. 10 is a block diagram illustrating the electrical configuration of a robot failure sign detection device according to a fifth embodiment. [Figure 15] FIG. 13 is a block diagram illustrating the electrical configuration of a robot failure sign detection device according to a sixth embodiment. [Figure 16] 10 is a display example of a trend management screen displayed on the display unit in an example of predicting failure time without using a judgment model. [Figure 17] 10 is a graph showing changes in the prediction line when the number of reference days is changed in an example in which failure time is predicted without using a determination model. [Figure 18] 10 is a graph showing a frequency band for calculating a frequency analysis partial integrated value. [Figure 19] 10 is a display example of a trend management screen displayed on the display unit in an example where the frequency analysis partial integrated value can be set. DETAILED DESCRIPTION OF THE INVENTION

[0014] Next, an embodiment of the present application will be described with reference to the drawings. Fig. 1 is a perspective view showing the configuration of a robot 1 according to an embodiment of the present application. Fig. 2 is a block diagram showing the electrical configuration of the robot 1 and a robot failure sign detection device 5.

[0015] The robot failure sign detection device 5 according to the present application is used to monitor the state of an industrial robot capable of reproducing predetermined operations. The robot failure sign detection device 5 is applied to, for example, a robot 1 as shown in FIG. 1. The robot 1 performs tasks such as painting, cleaning, welding, and transporting a workpiece. The robot 1 is realized, for example, by a vertical articulated robot.

[0016] The configuration of the robot 1 will be briefly described below with reference to FIGS. 1 and 2.

[0017] The robot 1 comprises a base member 10, an articulated arm 11, and a wrist 12. The base member 10 is fixed to the ground (for example, the floor of a factory). The articulated arm 11 has a plurality of joints. The wrist 12 is attached to the tip of the articulated arm 11. An end effector 13 is attached to the wrist 12 to perform work on a workpiece.

[0018] As shown in FIG. 2, the robot 1 includes an arm driving device 21.

[0019] These drive devices are composed of actuators, reducers, etc. The actuators are composed of, for example, servo motors. However, the configuration of the drive devices is not limited to the above. Each actuator is electrically connected to a controller 90. The actuators operate to reflect command values ​​input from the controller 90.

[0020] The driving force from each servo motor constituting the arm driving device 21 is transmitted via a reducer to each joint of the articulated arm 11, the base member 10, and the wrist 12. Each servo motor is fitted with an encoder (not shown) that detects its rotational position.

[0021] The robot 1 performs tasks by playing back the motions recorded through instruction. The controller 90 controls the actuators so that the robot 1 reproduces the series of motions previously taught by the instructor. The instructor can teach the robot 1 by operating a teaching pendant (not shown). A program for moving the robot 1 is generated by teaching the robot 1.

[0022] The controller 90 is configured as a known computer including, for example, a CPU, a ROM, a RAM, an auxiliary storage device, etc. The auxiliary storage device is configured as, for example, an HDD, an SSD, etc. The auxiliary storage device stores programs for operating the robot 1, etc.

[0023] 1, the robot failure sign detection device 5 is connected to a controller 90. The robot failure sign detection device 5 acquires, via the controller 90, the transition of the current value of the current flowing through the actuator (servo motor), etc.

[0024] If an abnormality occurs in the servo motor or the reducer connected to it, the current value of the servo motor is expected to fluctuate accordingly. Therefore, this current value can be said to reflect the state of the robot 1. The transition of the current value can be expressed by repeatedly obtaining the current value at short time intervals and arranging many current values ​​in a time series. Hereinafter, data in which current values ​​are arranged in a time series may be referred to as current value time series data (behavior time series data).

[0025] The robot failure sign detection device 5 can determine whether or not there is an abnormality in the robot 1 by monitoring the acquired current value time series data. In this embodiment, the robot failure sign detection device 5 determines whether or not there is an abnormality, mainly targeting the servo motors and reducers of each joint. Here, "abnormality" includes cases where some condition that does not lead to malfunction / inability to operate but is a precursor to such an abnormality occurs in the servo motor, reducer, or bearing.

[0026] As shown in FIG. 2, the robot failure sign detection device 5 includes a memory unit 50, a current value time series data acquisition unit (behavior time series data acquisition unit) 51, an evaluation value calculation unit 52, a representative evaluation value generation unit 53, a series processing unit 54, a judgment unit 55, an alarm generation unit 62, and a display unit 63.

[0027] The robot failure sign detection device 5 is configured as a known computer including a CPU, ROM, RAM, auxiliary storage device, etc. The auxiliary storage device is configured as, for example, an HDD, SSD, etc. The auxiliary storage device stores programs and the like for evaluating the state of the robot 1. Through cooperation of these hardware and software, the computer can operate as a memory unit 50, a current time series data acquisition unit 51, an evaluation value calculation unit 52, a representative evaluation value generation unit 53, a series processing unit 54, a judgment unit 55, an alarm generation unit 62, a display unit 63, etc.

[0028] The current value time series data acquiring unit 51 acquires the above-mentioned current value time series data. The current value time series data acquiring unit 51 acquires the current value time series data for all servo motors provided in the arm driving device 21 of the robot 1. The current value time series data is acquired individually for each of the multiple servo motors (in other words, multiple reducers) arranged in each part of the robot 1.

[0029] Here, the current value refers to the measured value of the current flowing through the servo motor, measured by a sensor. The signal from the sensor is digitized by an A / D converter (not shown). The sensor is provided in a servo driver (not shown) that controls the servo motor. However, a sensor may be provided separately from the servo driver for monitoring purposes. Alternatively, a current command value given to the servo motor by the servo driver may be used. The servo driver performs feedback control on the servo motor so that the current current value approaches the current command value. Therefore, for the purpose of detecting abnormalities in the servo motor or reducer, there is almost no difference between the current value and the current command value.

[0030] The magnitude of the torque of a servo motor is proportional to the magnitude of the current. Therefore, a torque value or torque command value may be used instead of a current value. Also, the deviation (rotational position deviation) between a target value for the rotational position of the servo motor and the actual rotational position obtained by the encoder may be used. Typically, a servo driver multiplies this deviation by a gain and provides the result to the servo motor as a current command value. Therefore, the transition of the rotational position deviation shows a similar trend to the transition of the current command value.

[0031] In this embodiment, the current value time series data is used to detect a sign of a failure. However, instead of the current value, time series data of a current command value, a torque value, a torque command value, or a rotational position deviation may be used.

[0032] The current value time series data acquisition unit 51 acquires current value time series data for each servo motor when the robot 1 performs the taught motion.

[0033] As an example, consider a case where the robot 1 is taught one operation and then operates in a factory every day, for example, from 9:00 to 17:00. The current time-series data acquisition unit 51 acquires current time-series data for each servo motor every time the robot 1 performs the taught operation. Because the robot 1 repeats the same operation, a large amount of current time-series data is obtained each day.

[0034] The period after the robot failure sign detection device 5 starts to be used can be divided into the period from day 1 to day N and the period from day N+1 onwards. Hereinafter, the period from day 1 to day N may be referred to as the initial operation period, and the period from day N+1 onwards may be referred to as the monitoring period. N can be determined as appropriate, and can be set to 30, for example. The current value time series data is acquired every day during both the initial operation period and the monitoring period.

[0035] The initial operation period may start at the same time as the robot is installed and begins to be used, but it may also be considered a break-in period of, for example, one month, and the initial operation period may start after the break-in period. After the initial operation period, the monitoring period may start after an appropriate interval period (for example, two months) has elapsed.

[0036] Current time-series data is obtained for each servo motor every time the robot 1 performs one operation. The timing at which the current time-series data acquiring unit 51 starts and ends acquiring the current time-series data can be determined appropriately based on the signal output by the controller 90.

[0037] The graph in Fig. 3 shows an example of the current value flowing through the servo motor of a certain joint when the robot 1 performs a playback operation. As shown in Fig. 3, before the playback operation program is executed, the current value of the servo motor is zero. At this time, electromagnetic brakes (not shown) are operating in each joint, so the posture of the articulated arm 11 and the like is maintained.

[0038] Next, the playback operation program for the robot 1 is started. This releases the brake, and at almost the same time, current begins to flow to the servo motor. At this point, the output shaft of the servo motor is controlled to stop. After a certain amount of time has passed, which is necessary for the angle of the output shaft of the servo motor to stabilize, the servo motor starts to rotate. This essentially starts the operation of the robot 1.

[0039] After the brake is released, the controller 90 outputs an acquisition start signal to the robot failure sign detection device 5 (and further to the current value time series data acquisition unit 51) shortly before the servo motor starts to rotate.

[0040] When the series of operations taught to the robot 1 are all completed, the servo motor is controlled to stop rotating. After the servo motor stops rotating and before the program ends, the controller 90 outputs an acquisition end signal to the current value time series data acquisition unit 51.

[0041] The storage unit 50 is configured, for example, from the auxiliary storage device described above. The storage unit 50 stores a robot failure sign detection program, current value time series data acquired by the current value time series data acquisition unit 51, and the like. The robot failure sign detection program realizes the robot failure sign detection method of this embodiment.

[0042] In this embodiment, the time series data is a chronological arrangement of a large number of current values ​​obtained by repeatedly detecting the current values ​​at short, fixed time intervals. The time interval (sampling interval) for detecting the current values ​​is, for example, several milliseconds. The time series data corresponds to the transition of the current value from the timing of the acquisition start signal to the timing of the acquisition end signal in the graph of FIG. 3.

[0043] The storage unit 50 stores time information indicating the time when the data was acquired in association with the time-series data.

[0044] As described above, the current time series data acquired by the current time series data acquisition unit 51 can be divided into data collected during the initial operation period of the robot 1 and data collected during the monitoring period. The current time series data acquired during the initial operation period is used to create learning data for constructing a determination model, which will be described later. The current time series data collected during the monitoring period is used to create determination data to be input to the determination model.

[0045] Appropriate filtering may be performed on the current time series data to remove noise, etc. Filtering of data is well known, and therefore a detailed description thereof will be omitted.

[0046] The evaluation value calculation unit 52 inputs each piece of current time series data acquired by the current time series data acquisition unit 51 into a predetermined function and calculates an evaluation value. This function extracts some feature of the current time series data. While the original current time series data consists of multiple scalar quantities (current values), the evaluation value is a single scalar quantity. Therefore, the processing performed by the evaluation value calculation unit 52 corresponds to information compression.

[0047] The evaluation value calculation unit 52 may use any function, for example, the root mean square, maximum value, value range (maximum value-minimum value), or frequency analysis integrated value.

[0048] As is well known, changes occur in the root mean square of the current value when the efficiency of the reducer decreases due to wear on the teeth, when the torque constant decreases due to demagnetization of the magnet of the servo motor, etc. Therefore, it is sometimes preferable to use the root mean square to determine whether there are signs of a failure.

[0049] Furthermore, when damage occurs to bearings, etc., whisker-like peaks tend to appear from time to time in the transition of the current value. Therefore, it may be preferable to use the maximum value or value range to determine whether or not there is a sign of failure.

[0050] The frequency analysis integrated value is proposed as one method of frequency analysis. By performing frequency analysis, an amplitude spectrum, a power spectrum, or a power spectral density can be obtained. Hereinafter, the amplitude spectrum, the power spectrum, or the power spectral density may be referred to as the frequency spectrum or simply the spectrum. The frequency analysis integrated value is, for example, the sum of the amplitude spectrum, the power spectrum, or the power spectral density up to several tens of Hz obtained by frequency analysis. The frequency analysis integrated value may also be obtained using an average value instead of the sum. The amplitude spectrum and the power spectrum are obtained using a known method, which will be described later. It may be appropriate to use the frequency analysis integrated value to determine whether or not a robot is showing signs of failure, such as a tendency for vibration. There are various possible causes for a tendency for vibration, but one example is increased lost motion due to wear on the reducer.

[0051] If the natural frequency of the system of robot 1 is 8 Hz, the 8 Hz component will become large due to resonance, but the example of the integrated frequency analysis value above does not simply detect the 8 Hz peak. The total value of the amplitude spectrum, power spectrum, or power spectral density up to several tens of Hz is evaluated. Over time, in addition to the original main mode vibration, vibrations due to the deterioration of various parts often occur. By observing a wide range of frequencies, the range of detection for signs of failure can be expanded.

[0052] The frequency analysis integrated value will be explained in more detail below. The evaluation value calculation unit 52 obtains a Fourier spectrum (complex number) by performing an FFT (Fast Fourier Transform) on the current value time series data. The value obtained by squaring the Fourier spectrum is the power spectrum. Phase information is lost when converted to a power spectrum. The value obtained by taking the square root of the power spectrum is the amplitude spectrum. The amplitude spectrum has an effective value and a peak value, and either value can be used to determine whether or not there is a fault sign. The power spectral density function is also called a PSD function (Power Spectral Density Function). The power spectral density function is a spectral function that expresses the power value per unit frequency width (1 Hz width) independent of the frequency resolution of the FFT.

[0053] The upper limit of the frequency that can be detected by FFT is half the sampling frequency. Half the sampling frequency is called the Nyquist frequency. In this embodiment, the sampling frequency corresponds to the reciprocal of the longer of the control periods of the robot 1 and the robot failure sign detection device 5. For example, if the longer control period is 4 ms, the Nyquist frequency is 125 Hz, since 1 / 0.004 / 2 = 125. The FFT performed in this embodiment targets values ​​below the Nyquist frequency. However, depending on the hardware configuration or software algorithm of the device executing the FFT, it is preferable to use a value multiplied by a margin factor as the upper limit to reduce or eliminate aliasing. That is, if the Nyquist frequency is written as f9 and the margin factor is written as k, the upper limit of the FFT can also be expressed as k × f9. k is, for example, a value ranging from 75% to 100%.

[0054] When the behavior of the robot 1 is measured from the current value or current command value of the actuator and FFT is performed as in this embodiment, there is an upper limit based on the Nyquist frequency, etc. The software that executes the FFT may be regulated so as not to exceed the Nyquist frequency.

[0055] On the other hand, the robot 1 normally has a natural frequency of around 10 Hz. Normally, if the first-order natural frequency is 10 Hz, signals of a certain strength can be detected up to the third or fourth order components. Therefore, if the natural frequency is f0 and signals up to the fourth order are to be used, for example, the upper limit should be set to 4.5 x f0 = 45 Hz. This means that the fourth order signals will be acquired but the fifth order signals will not be acquired, allowing for a more accurate understanding of the vibration components.

[0056] In addition to the natural frequency, there are also vibrations that occur due to faults. For example, vibrations other than the natural frequency can occur due to damage to the motor bearings, damage to the teeth of the input shaft gear and revolving gear, and damage to the eccentric shaft bearing. Among these, damage to the motor bearings has a relatively high frequency, and vibrations of around 25 Hz, for example, can occur. Note that the value of the frequency also varies depending on whether the damage occurs in the inner ring, outer ring, or rolling element of the bearing, and on the motor rotation speed. Taking all of the above into consideration, it is preferable to set the upper limit at around 50 Hz, for example. Alternatively, the Nyquist frequency may be set as the upper limit.

[0057] Furthermore, if emphasis is placed on the main natural frequency, the spectrum can be searched in order from 0 Hz, the first peak value being the primary natural frequency f0, and the results can be tallied up to 4.5 times that, or 4.5 x f0. In this case, the reference frequency to be searched can be registered as the initial frequency, and a search can be performed from 0.5 to 1.5 times the initial frequency, for example. If the frequency characteristics are stable, a search can be performed from 0.2 to 1.8 times. If this margin is expressed as Δf1 and Δf2, and the initial frequency of the natural frequency to be searched is expressed as f01, the search will be performed from f01 - Δf1 to f01 + Δf2.

[0058] In the articulated robot 1 of this embodiment, the entire articulated arm 11 can be rotated relative to the base member 10, for example, with the vertical direction as the center of rotation. This rotation axis is referred to as the rotation axis JT1. The inertia around the rotation axis JT1 changes significantly depending on the posture of the articulated arm 11. As a result, the natural frequency f0 may vary between 10 Hz and 20 Hz. In the case of such an axis, it may not be possible to search for the first-order natural frequency f0 as described above. In this case, a "deemed" natural frequency f02 can be registered as, for example, 12 Hz, and up to 4.5 times that frequency can be calculated. From the above, the upper and lower limits of the integrated frequency can be summarized as follows: Furthermore, any combination of the upper and lower limits described below may be used.

[0059] As the upper limit, for example, the following values ​​can be used. (1) Fixed value (2) (n+0.5) times the first natural frequency (3) (n+0.5) times the assumed frequency The fixed value may be, for example, 50 Hz. However, the fixed value must be equal to or less than the Nyquist frequency. The fixed value may be the Nyquist frequency or may be a value obtained by multiplying the Nyquist frequency by a margin factor. n is, for example, 4, but may be any value, such as 1, 2, or 3. Also, instead of 0.5, another value greater than or equal to 0 and less than 1 may be used.

[0060] As the lower limit, for example, the following values ​​can be used. (1) Fixed value (2)0 (3) α times the natural frequency Here, the fixed value may be, for example, 6 Hz. Also, 0 is called a DC component, and is a quantity that does not involve vibration, and corresponds to the load torque of each axis of the robot 1. Therefore, when evaluating the load torque as well, 0 is used as the lower limit, and when evaluating only the vibration component, α times the natural frequency can be used as the lower limit. α is, for example, 0.5, but α may be any value between 0.2 and 0.8. Note that when evaluating only the vicinity of the natural frequency, the lower limit α may be 0.5, for example, and the upper limit may be 1.5 times the natural frequency, for example.

[0061] As described above, different functions result in different fault signs that are easier to detect. Since the output values ​​of the functions are essentially input into the judgment model described below and machine learning is performed, the judgment model also inherits the properties of the functions. Therefore, the judgment model may be trained and evaluated for each of the evaluation values ​​obtained using multiple different functions.

[0062] In the above example, the frequency analysis integrated value is used as a single function for evaluation. In contrast, in the example shown below, the frequency analysis integrated value is divided into multiple frequency bands, and the total value is calculated for each range. To distinguish it from the frequency analysis integrated value, the integrated value of the frequency spectrum for each frequency band is called the "frequency analysis partial integrated value." It is sometimes abbreviated to "partial integrated value." The frequency analysis partial integrated value may also be calculated using an average value instead of the total value.

[0063] The partial integration value will be explained with reference to Figure 18. The vertical axis in Figure 18 is the amplitude of the spectrum, and the horizontal axis is the frequency. In this example, the lower limit of the frequency is 0 Hz and the upper limit is 50 Hz. Furthermore, values ​​of 10, 20, 30, and 40 are set to divide the range from 0 Hz to 50 Hz. By dividing the spectrum using these values, the following frequency bands are obtained. Frequency bands: 0-10, 10-20, 20-30, 30-40, 40-50 Each frequency band may or may not include both end frequencies (lower and upper limit frequencies for determining the frequency band). Also, the end frequencies may be included in two adjacent frequency bands.

[0064] Here, we will explain the vibrations generated by robot 1 using the 10-20 frequency band indicated by the bold frame in Figure 18 and the smaller 0-10 frequency band as examples. For example, the 0-10 frequency band includes a spectrum of approximately 8 Hz, which is the natural frequency of the robot arm system. Large amplitudes also exist at 4, 2, and 1 Hz, but most of these frequencies are frequency components of the trajectory of the robot's movement. The frequency components of these trajectories hardly change when the robot program that determines the movement is the same. The 10-20 frequency band includes the second harmonic of the natural frequency of approximately 16 Hz. The 10-20 frequency band also includes components other than harmonics of the natural frequency.

[0065] The spectrum of the natural frequency and its harmonics can become larger due to changes in the spring constant of the reducer, which determines the natural frequency, or an increase in the amount of lost motion. Conversely, the spectrum of the natural frequency and its harmonics can also increase due to deterioration of other components (such as bearings attached to the motor shaft, etc.). Furthermore, in either case, not only does the amplitude change (usually an increase), but the frequency also often changes. As a result of deterioration, the natural frequency can decrease, and if that frequency is near a divided frequency band, it may shift to a lower frequency band. Harmonics of the natural frequency are present even in bands above 20 Hz, but their intensity decreases, and the proportion of other vibration components increases.

[0066] As explained above, the partial integrated values ​​obtained by dividing the data into multiple parts can be used in the determination model described below. The partial integrated values ​​can also be used in a lifespan prediction or failure sign determination system that simultaneously operates multiple trend management devices.

[0067] Note that the divisions do not have to be continuous as shown in Figure 18, and there may be intervals between the frequency bands, for example, 0-10, 20-30, and 40-50. Alternatively, the frequency bands may overlap, such as 0-10, 0-20, and 0-30. Since the partial integrated value is used for trend management, the partial integrated value itself has no meaning; changes in the partial integrated value are evaluated. Therefore, it is not necessary to include all frequencies without omissions or overlaps; it is sufficient to select frequency bands that are easy to manage trends. Of course, as shown in Figure 18, it is also possible to include all frequencies without omissions or overlaps.

[0068] The representative evaluation value generation unit 53 calculates a representative evaluation value from the multiple evaluation values ​​obtained by the evaluation value calculation unit 52. As mentioned above, one evaluation value corresponds to compressed information of the current value time-series data, but the representative evaluation value can be thought of as information obtained by further compressing multiple evaluation values.

[0069] For example, consider a case where the robot 1 operates from 9:00 to 17:00 on a certain day. During that period, the robot 1 repeats the same operation many times. Therefore, the current time-series data acquisition unit 51 obtains a large number of time-series current value data, and the evaluation value calculation unit 52 obtains an evaluation value for each time-series current value data. The representative evaluation value generation unit 53 generates one evaluation value that represents that day from the large number of evaluation values.

[0070] For example, the room temperature in a factory may change between 9:00 and 17:00 due to the influence of the outside temperature. Alternatively, the grease applied to the joints of the robot 1 is low when the robot 1 first starts operating, but as time passes, its temperature rises and, as a result, its viscosity decreases. To suppress these effects, the representative evaluation value generation unit 53 may select a median value from the many evaluation values ​​obtained between 9:00 and 17:00 and use this as the representative evaluation value. The representative evaluation value generated by the representative evaluation value generation unit 53 is stored in the storage unit 50.

[0071] The representative evaluation value is not limited to the median value, but may be, for example, the average value.

[0072] Hereinafter, the time range to which multiple evaluation values ​​represented by one representative evaluation value belong is called the "data collection unit period." In the above example, one representative evaluation value represents one day's worth of current time-series data, so the data collection unit period is one day.

[0073] For example, in a factory, multiple shifts (e.g., a two-shift system) may be set up in one day. One shift can also be referred to as one work cycle or one operation cycle of the robot 1. In this case, the data collection unit period may be one shift, but it is preferable to set it to one day including two shifts. By setting one day including day and night as the data collection unit period, it is possible to prevent a decrease in the accuracy of detecting signs of failure due to the influence of fluctuation cycles of the outside air temperature. The data collection unit period may be multiple days (e.g., two days, three days, or one week).

[0074] The series processing unit 54 generates a series by arranging the representative evaluation values ​​stored in the storage unit 50. The series of representative evaluation values ​​generated by the series processing unit 54 is input to the determination unit 55.

[0075] The determination unit 55 uses the representative evaluation value generated by the representative evaluation value generation unit 53 to determine whether or not there is a sign of a failure in the robot 1. The determination unit 55 includes a learning unit 57, a probability output unit 58, and a conversion unit 61.

[0076] As described above, by using the representative evaluation value instead of the evaluation value, it is possible to eliminate the influence of temperature changes over a day, for example. As a result, this may have the effect of compressing the evaluation value. On the other hand, if the evaluation method is trend management or the evaluation value is temperature compensated using some method, it is possible to evaluate all the obtained evaluation values ​​without using a representative evaluation value.

[0077] During the initial operation period of the robot 1, the learning unit 57 learns N representative evaluation values ​​generated in N data unit periods, thereby creating a determination model, which is a machine learning model.

[0078] In the determination unit 55 of this embodiment, a hidden Markov model is used as a determination model.

[0079] Here, we will briefly explain the hidden Markov model. The hidden Markov model is a type of probabilistic model that realizes statistical modeling for sequential data. When sequential data is given, the hidden Markov model can calculate the probability that that sequential data will appear.

[0080] In the hidden Markov model, when sequential data is observed, it is assumed that there is a sequential state behind it. The state cannot be observed, only the data can be observed.

[0081] Below, we will explain the hidden Markov model. There are three parameters that define the hidden Markov model: transition probability A, output probability B, and initial probability Π. Here, the number of states is R. The initial probability Π is the same symbol as Π, which represents a sum in equation (1) described later, but its meaning is different.

[0082] The transition probability A is a ij (A={a ij}). a ij represents the probability that state i at time (t-1) becomes state j at time t, where i and j are integers between 1 and R.

[0083] The output probability B is b jk (B={b jk}). b jk is the kth observed signal v in state j. k represents the probability distribution of the output. In this embodiment, the output is assumed to be a continuous value, and the observed signal v k The probability that the output is expressed by a normal distribution. In this embodiment, a single normal distribution is used, but a mixed normal distribution may be used instead.

[0084] The initial probability Π is π j is a set of (Π={π j}). π jrepresents the probability of being in state j at time t = 0. In a hidden Markov model, the state is hidden, but the initial value is arbitrarily tentatively determined.

[0085] Hidden Markov models are roughly divided into ergodic hidden Markov models and left-to-right hidden Markov models.

[0086] An ergodic hidden Markov model is shown in Figure 4(a). An ergodic hidden Markov model contains multiple states. Figure 4(a) shows an example with three states. Each state can transition to any state, including itself.

[0087] Ergodic hidden Markov models are ergodic. Ergodicity means that [A] any state can be reached from any other state, [B] there is no periodicity, and [C] the number of states is finite. When a hidden Markov model is ergodic, the ensemble average and the time average are the same.

[0088] A Left-to-Right Hidden Markov Model is shown in Figure 4(b). A Left-to-Right Hidden Markov Model contains multiple states. Figure 4(b) shows an example with four states. In a Left-to-Right Hidden Markov Model, state transitions are always unidirectional, so once a state transitions to another state, it cannot return to the state before the transition.

[0089] The Left-to-Right Hidden Markov Model does not have ergodicity. The Left-to-Right Hidden Markov Model has a restriction that prevents reverse transitions. Because of this restriction on state transitions, the Left-to-Right Hidden Markov Model has the advantage of reducing the amount of calculation required, making it suitable for evaluating time series data.

[0090] In this embodiment, either an Ergodic Hidden Markov Model or a Left-to-Right Hidden Markov Model can be used as the determination model. The number of states of the Hidden Markov Model is any number equal to or greater than two and can be determined appropriately.

[0091] The learning unit 57 inputs the above-mentioned series of N representative evaluation values ​​into the determination model, and updates the above-mentioned transition probability, output probability, and initial probability parameters so that the determination model will have a higher probability of generating the series of representative evaluation values. At this time, the well-known EM (Expectation Maximization) method and Baum-Welch algorithm are used.

[0092] When an Ergodic Hidden Markov Model is used, the order in which the representative evaluation values ​​are input to the learning model may or may not be fixed. When a Left-to-Right Hidden Markov Model is used, it is preferable that the order in which the representative evaluation values ​​are input is fixed in chronological order. The sequence processing unit 54 generates a sequence consisting of N representative evaluation values ​​in accordance with conditions and outputs it to the determination unit 55.

[0093] The probability output unit 58 inputs a sequence including the representative evaluation value obtained during the monitoring period into the created determination model, and obtains the probability that the sequence will appear.

[0094] By inputting a sequence of representative evaluation values, the hidden Markov model can calculate and estimate the probability that this sequence will appear. The calculated probability can also be thought of as a numerical representation (likelihood) of the plausibility of the sequence input to the model. If the probability output from the model is low, it can be said that there is a high probability that the sequence of representative evaluation values ​​input to the model is different from normal. Therefore, this probability reflects the specificity. The specificity can also be rephrased as the degree of anomaly. The more similar the input sequence is to the learned sequence, the higher the probability the hidden Markov model outputs. Therefore, obtaining the probability is essentially the same as obtaining the similarity between the two sequences. The probability output unit 58 outputs the obtained probability to the conversion unit 61.

[0095] The conversion unit 61 performs logarithmic conversion on the probability obtained from the probability output unit 58. The obtained value (log likelihood) is output to the warning generation unit 62 and the display unit 63.

[0096] The warning generation unit 62 issues a warning of a failure sign when the log likelihood input from the conversion unit 61 is outside a predetermined range. In this embodiment, the warning is realized by the warning generation unit 62 controlling the display of the display unit 63.

[0097] The display unit 63 can display graphs such as those shown in FIG. 8. The display unit 63 is configured with a display device such as a liquid crystal display. The operator monitors whether the log likelihood of the series of representative evaluation values ​​deviates from the usual trend. The operator can use this information to appropriately plan future maintenance. In addition, the display unit 63 can switch the axis to be displayed and output a warning in the form of a message, for example, in response to a signal from the warning generation unit 62.

[0098] Next, a robot failure sign detection method used by the robot failure sign detection device 5 of this embodiment will be described along the data flow shown by the thin line in Fig. 2. In the following description, the data collection unit period is set to one day and N is set to 30.

[0099] The current value time series data acquisition unit 51 of the robot failure sign detection device 5 acquires current value time series data for each of multiple robot movements played back in a given day via the current value time series data acquisition unit 51 (behavior time series data acquisition process).

[0100] Thereafter, the evaluation value calculation unit 52 calculates an evaluation value using a function such as the root mean square from each piece of current time series data acquired by the current time series data acquisition unit 51 (evaluation value calculation step). The number of evaluation values ​​obtained is equal to the number of pieces of current time series data.

[0101] Next, the representative evaluation value generating unit 53 generates a representative evaluation value that represents the multiple evaluation values ​​obtained within a day (representative evaluation value generating step). The representative evaluation value can be the median value of the multiple evaluation values ​​obtained within a day.

[0102] The above process is repeated every day, resulting in 30 representative evaluation values ​​being obtained at the end of N data collection unit periods (30 days) that constitute the initial operation period.

[0103] Once the 30 representative evaluation values ​​are obtained, the series processing unit 54 generates a series by chronologically arranging the representative evaluation values ​​from the first day to the 30th day (series processing step). The series processing unit 54 outputs the generated series of representative evaluation values ​​to the determination unit 55. Hereinafter, this series may be referred to as the initial series. The initial series is shown in FIG. 6(a). In FIG. 6, one rectangle represents one representative evaluation value. The number inside the rectangle indicates the day on which the representative evaluation value was obtained.

[0104] The learning unit 57 of the judgment unit 55 inputs this initial sequence into a machine learning model for learning (model creation process). A sequence of representative evaluation values ​​is input into the judgment model, and the parameters of the judgment model are modified so that the occurrence probability of the sequence increases. Specifically, the parameters of the judgment model are the transition probability A, output probability B, and initial probability Π described above. This process corresponds to one learning session, and learning is repeated multiple times. Repetition of this process corresponds to the training phase of machine learning.

[0105] The number of learning times can be set to, for example, 100. However, learning can be terminated if convergence of the parameters of the determination model is confirmed even before learning reaches 100 times. After learning is completed, the parameters of the determination model are stored in a determination model storage unit 59 provided in the determination unit 55.

[0106] The N+1th day and onwards constitute the monitoring period. During the monitoring period, the series processing unit 54 creates a series of representative evaluation values ​​whose specificity will be evaluated by the determination unit 55 (series processing step). The series processing unit 54 outputs the generated series of representative evaluation values ​​to the determination unit 55. Hereinafter, this series may be referred to as a determination series. In this embodiment, the determination series corresponds to determination data. The determination series generated by the series processing unit 54 changes every time the representative evaluation value generation unit 53 generates a new representative evaluation value, i.e., every day. The determination series includes at least one representative evaluation value from the N+1th day onwards.

[0107] The processing for the judgment series will be explained in detail below. When the representative evaluation value for the (N+1)th data collection unit period (the 31st day) is obtained, the series processing unit 54 initializes the judgment series with the initial series. This initialization processing is performed only the first time (the 31st day). Next, the series processing unit 54 adds the most recent representative evaluation value obtained to the end of the judgment series. This addition brings the number of representative evaluation values ​​constituting the judgment series to 31, so the series processing unit 54 deletes the oldest representative evaluation value located at the beginning of the series. The judgment series on the 31st day is shown in Figure 6(b).

[0108] This update process is performed in the same way for the 32nd day and thereafter. In this way, the series processing unit 54 deletes the oldest representative evaluation value from the previous judgment series, moves the second and subsequent representative evaluation values ​​one position toward the beginning of the series, and adds the latest representative evaluation value to the end, thereby creating a series of representative evaluation values ​​for each day. The judgment series on the 32nd day is shown in FIG. 6(c), and the judgment series on the 33rd day is shown in FIG. 6(d). The representative evaluation values ​​of the judgment series are replaced one by one with each update process.

[0109] The probability output unit 58 inputs the judgment sequence generated by the sequence processing unit 54 into the judgment model stored in the judgment model storage unit 59, and calculates the probability that the sequence will appear. This process corresponds to the evaluation phase of machine learning. The probability output unit 58 outputs the obtained probability to the conversion unit 61.

[0110] The conversion unit 61 converts the probability output from the probability output unit 58 into a logarithmic value to obtain a logarithmic likelihood. Logarithmic conversion makes it easier to handle numerical values.

[0111] The warning generation unit 62 determines whether or not there is a sign of failure by checking whether the obtained log likelihood is within a predetermined range (failure sign determination step). Specifically, the warning generation unit 62 compares the log likelihood with a predetermined threshold. If the log likelihood falls below the predetermined threshold, the warning generation unit 62 issues a warning, for example, by displaying on the display unit 63 that a sign of failure has been detected (warning generation step).

[0112] Next, the effects of this embodiment using a hidden Markov model will be described.

[0113] The graph in Figure 7 shows the change in the root mean square (sometimes called I2) of the current value of a certain servo motor since the robot started operating. The horizontal axis is time, and the vertical axis is the root mean square. In this example graph, we can see that the root mean square increases slightly from around the triangle mark. However, because this increase is small, it is difficult to determine whether or not there are signs of a malfunction. The thick line in Figure 7 represents the value obtained by calculating the average over the first 30 days and multiplying the obtained value by 1.1. If this thick line is used as the threshold for determining an abnormality, only one point will be determined to be abnormal.

[0114] The graph in Figure 8 shows the results of evaluating the root mean square using a hidden Markov model. The horizontal axis is time, and the vertical axis is log-likelihood. In the example in Figure 8, the Ergodic hidden Markov model with two states shown in Figure 5 is used.

[0115] As shown in Figure 8, the log likelihood began to change around December 16th. Signs of an abnormality are clearly visible in the slope-like transition of multiple points (rather than appearing as a single outlier point as in Figure 7). This is thought to be because the proportion of anomalous data in the judgment sequence tends to increase cumulatively each time the judgment sequence is updated.

[0116] This property allows the operator to intuitively and easily grasp signs of abnormality.

[0117] As described above, the robot failure sign detection device 5 of this embodiment includes a current time series data acquisition unit 51, an evaluation value calculation unit 52, a representative evaluation value generation unit 53, a series processing unit 54, and a judgment unit 55. The current time series data acquisition unit 51 performs processing for acquiring current time series data of the drive currents of the joints of the robot 1 from the robot's operation for each data collection unit period. The evaluation value calculation unit 52 calculates an evaluation value for the current time series data acquired by the current time series data acquisition unit 51. The representative evaluation value generation unit 53 generates a representative evaluation value representing the evaluation values ​​from the evaluation values ​​obtained by the evaluation value calculation unit 52 for each data collection unit period. The series processing unit 54 generates a series of representative evaluation values. The judgment unit 55 creates a judgment model based on the initial series, which is the series generated by the series processing unit 54, at the beginning of operation of the robot 1. After the initial stage of operation, the judgment unit 55 inputs a judgment series including a representative evaluation value based on the robot's operation after the initial stage of operation into the created judgment model to obtain the specificity of the judgment series.

[0118] This makes it possible to effectively detect signs of a malfunction in the robot 1. Therefore, it becomes possible to perform maintenance on the robot 1 before a malfunction occurs.

[0119] Furthermore, in this embodiment, after the initial stage of operation, the judgment unit 55 inputs a judgment series composed of multiple representative evaluation values ​​as judgment data to the judgment model for each data collection unit period, and obtains specificity based on the output of the judgment model. To generate the judgment series, the series processing unit 54 initializes the judgment series with the initial series and then updates the judgment series each time a representative evaluation value is generated. The update process includes adding a representative evaluation value for the data collection unit period to the judgment series before updating.

[0120] This makes it easier to detect signs of a malfunction in the robot 1 from the transition of the specificity.

[0121] Furthermore, in the robot failure sign detection device 5 of this embodiment, the judgment model is a hidden Markov model that has been trained on an initial sequence.

[0122] This makes it possible to detect signs of a failure in the robot 1 with high accuracy using a model that can easily handle sequential data.

[0123] In the robot failure sign detection device 5 of this embodiment, the hidden Markov model is an Ergodic hidden Markov model. The number of states that the Ergodic hidden Markov model has is two.

[0124] This allows the model configuration to be simplified.

[0125] Furthermore, in the robot failure sign detection device 5 of this embodiment, the hidden Markov model may be a left-to-right hidden Markov model that is trained on an initial sequence. In this case, the initial sequence and the judgment sequence are configured as sequences in which N representative evaluation values ​​are arranged in chronological order.

[0126] This makes it possible to use a model that is good at handling time-series data to detect with high accuracy signs of a failure in the robot 1. In addition, the amount of calculation can be reduced by the amount of constraints on state transitions in the model.

[0127] Furthermore, in the robot failure sign detection device 5 of this embodiment, the root mean square is calculated as the evaluation value. However, the maximum value, the value range, or the frequency analysis integrated value may also be calculated as the evaluation value.

[0128] This allows for a good evaluation of changes in the current value time series data.

[0129] Furthermore, in the robot failure sign detection device 5 of this embodiment, the evaluation value calculation unit 52 performs frequency analysis on each of the behavior time series data acquired by the current value time series data acquisition unit 51 to obtain a frequency spectrum, and calculates multiple representative evaluation values ​​that are partial sums of the frequency spectra for multiple predetermined frequency bands. The representative evaluation value generation unit 53 generates multiple representative evaluation values ​​for each data collection unit period. The judgment model is a model that can input a multi-dimensional series as an initial series.

[0130] This makes it possible to determine whether or not there is a sign of a failure in the robot 1 based on the frequency spectrum obtained through frequency analysis.

[0131] Furthermore, in the robot failure sign detection device 5 of this embodiment, the data collection unit period is determined to be an integer multiple of the change cycle of the environmental temperature, for example, an integer multiple of one day.

[0132] This effectively eliminates the influence of periodic changes in the environmental temperature, thereby improving the accuracy of detecting signs of failure.

[0133] Furthermore, in the robot failure sign detection device 5 of this embodiment, the data collection unit period is determined to be an integer multiple of the work cycle or the operation cycle of the robot 1.

[0134] This makes it possible to substantially eliminate the influence of the phenomenon that the viscosity of the grease is high at the beginning of a work cycle or an operating cycle of the robot 1. Therefore, the accuracy of detecting signs of failure is good.

[0135] The robot failure sign detection device 5 of this embodiment also includes a display unit 63 that displays the uniqueness output by the determination unit 55.

[0136] This allows the signs of a malfunction in the robot 1 to be easily communicated to those around.

[0137] In this embodiment, the determination unit 55 outputs the specificity in a logarithmic transformed form.

[0138] This makes it easier to numerically handle the output of the determination unit 55. For example, it is possible to prevent the output of the graph in FIG.

[0139] The robot failure sign detection device 5 of this embodiment also includes an alarm generation unit 62. The alarm generation unit 62 issues an alarm of a failure sign when the uniqueness output by the determination unit 55 falls outside a predetermined range.

[0140] This allows the signs of a malfunction to be clearly communicated to those around.

[0141] Next, a second embodiment will be described. In the descriptions of the second and subsequent embodiments, the same or similar members as those in the previous embodiment will be denoted by the same reference numerals in the drawings, and descriptions thereof may be omitted.

[0142] In this embodiment, a model based on a normal distribution is used as the determination model instead of a hidden Markov model.

[0143] 9 is a block diagram of this embodiment. A determination unit 55 of this embodiment includes a determination model creation unit 67 instead of the learning unit 57 of the first embodiment. Other parts are substantially similar to those of the first embodiment.

[0144] The judgment model used in this embodiment will be described below. An initial sequence C consisting of N representative evaluation values ​​is input to the judgment model generation unit 67. 1(1) ,C 1(2) ,···,C 1(N) Consider the case where the following is input: The determination model creation unit 67 calculates the mean μ and standard deviation σ for the N representative evaluation values ​​that make up these initial series.

[0145] A determination model f is constructed in the determination unit 55. This determination model f is expressed by the following equation (1).

number

[0146] The above formula (1) includes the formula for a well-known normal distribution. That is, the judgment model f in this embodiment corresponds to a normal distribution based on the mean μ and standard deviation σ of the N representative evaluation values ​​in the initial sequence multiplied by the number Q of representative evaluation values ​​in the judgment sequence.

[0147] It has been confirmed through experiments by the inventors that the output when this determination model f is used also shows a transition substantially similar to that of the graph of Fig. 8 in the first embodiment. That is, the output of the determination model f can be treated as the probability (in other words, specificity) of the appearance of a sequence consisting of Q representative evaluation values.

[0148] In this embodiment, a model can be obtained by calculating the mean μ and standard deviation σ of the initial series. The mean μ and standard deviation σ calculated by the judgment model creation unit 67 are stored in the judgment model storage unit 59 as parameters that define the judgment model.

[0149] The judgment model f shown in equation (1) uses the multiplication of a normal distribution, but the multiplication or sum of the logarithm of a normal distribution may also be used. The sum of a normal distribution may also be used. According to experiments by the inventors, when the sum is used (however, without logarithmic transformation), the transition of the output of the judgment model f tends to fluctuate somewhat more roughly than when using the multiplication, but it is still possible to detect signs of a fault.

[0150] As described above, in the robot failure sign detection device 5 of this embodiment, the judgment model is formed by multiplying or summing Q times a normal distribution based on the standard deviation σ and mean value calculated from the N representative evaluation values ​​constituting the initial series. Furthermore, logarithmic conversion may be performed on the intermediate or final values.

[0151] This allows the determination model to be created with a small amount of calculation.

[0152] Next, a third embodiment will be described. Fig. 10 is a conceptual diagram showing the DTW method.

[0153] The robot failure sign detection device 5 of this embodiment is characterized by the calculation of the evaluation value performed by the evaluation value calculation unit 52. The evaluation value calculation unit 52 calculates a DTW distance or the like based on the DTW method from the current value time series data acquired by the current value time series data acquisition unit 51, and uses this as an evaluation value. DTW is an abbreviation for Dynamic Time Warping. In this embodiment, the parts other than the processing performed by the evaluation value calculation unit 52 are substantially similar to those of the first embodiment described above.

[0154] Here, we will briefly explain the DTW method. The DTW method is used to calculate the degree of similarity between two pieces of time series data. A major feature of the DTW method is that it allows for nonlinear expansion and contraction of the time series data along the time axis when calculating the similarity. This allows the DTW method to obtain results that are close to human intuition regarding the similarity of time series data.

[0155] The evaluation value calculation unit 52 outputs the DTW distance (dissimilarity) indicating the magnitude of the difference between the reference current value time series data (reference behavior time series data) and the comparison current value time series data (comparison behavior time series data) as an evaluation value of the current value time series data to be processed.

[0156] As the reference current time series data, for example, current time series data acquired by the current time series data acquiring unit 51 when the robot 1 is operated on a trial basis at the start of use of the robot failure sign detection device 5 is used. As the comparison current time series data, current time series data acquired by the current time series data acquiring unit 51 on or after the first day after the robot failure sign detection device 5 is started to be used. The reference data may be acquired on the first day, for example.

[0157] The principle of the DTW method will be explained using Figure 10. Multiple (s) current values ​​contained in the reference data are arranged in chronological order along a first axis extending horizontally. Multiple (p) current values ​​contained in the comparison data are arranged in chronological order along a second axis extending vertically.

[0158] Next, s × p cells are defined in a matrix on a plane defined by the vertical and horizontal axes. Each cell (l, m) represents the correspondence between the l-th current value in the reference data and the m-th current value in the comparison data, where 1≦l≦s and 1≦m≦p.

[0159] Each cell (l, m) is associated with a numerical value that represents the difference between the l-th current value in the reference data and the m-th current value in the comparison data. In this embodiment, each cell is associated with and stored with the absolute value of the difference between the l-th current value and the m-th current value.

[0160] The evaluation value calculation unit 52 finds a warping path (route) from the start cell located in the lower left corner of the matrix in FIG. 10 to the end cell located in the upper right corner.

[0161] The starting cell (1,1) corresponds to associating the current value at the earliest timing (i.e., the first) in the time series among the s current values ​​in the reference data with the current value at the earliest timing (i.e., the first) in the time series among the p current values ​​in the comparison data.

[0162] The end cell (s, p) corresponds to associating the current value at the last timing (i.e., the sth) in the time series among the s current values ​​in the reference data with the current value at the last timing (i.e., the pth) in the time series among the p current values ​​in the comparison data.

[0163] In the s × p matrix constructed as above, consider a path from the start cell to the end cell according to the following rules [1] and [2]. [1] You can only move to adjacent cells vertically, horizontally, or diagonally. [2] You cannot move backward in time in the reference data or in the comparison data.

[0164] A series of such cells is called a path or warping path. The warping path indicates how s current values ​​in the reference data correspond to p current values ​​in the comparison data. From another perspective, the warping path indicates how two time series data are stretched or compressed along the time axis.

[0165] There are multiple possible warping paths from the start cell to the end cell. The evaluation value calculation unit 52 finds the warping path among the possible warping paths that minimizes the sum of numerical values ​​representing the differences associated with the passing cells (in this embodiment, the absolute value of the difference between the l-th current value and the m-th current value). Hereinafter, this warping path may be referred to as the optimal warping path. Furthermore, the sum of the values ​​in each square on this optimal warping path may be referred to as the DTW distance.

[0166] The average DTW distance can be calculated by dividing the DTW distance by the number of squares passed. The average DTW distance may also be calculated by dividing the DTW distance by the number of elements s or p of the time series data. The average DTW distance can be used as the evaluation value instead of the DTW distance.

[0167] When s and p are large, a huge number of warping paths are possible. Therefore, if all possible warping paths were considered, the amount of calculation required to find the optimal warping path would increase explosively. To solve this problem, the evaluation value calculation unit 52 of this embodiment finds the optimal warping path using a DP matching method (dynamic programming). DP is an abbreviation for Dynamic Programming. The DP matching method is well known, so a description thereof will be omitted.

[0168] The DTW distance or the average DTW distance value is calculated by the evaluation value calculation unit 52 for each piece of current time series data acquired by the current time series data acquisition unit 51. The representative evaluation value generated by the representative evaluation value generation unit 53 is, for example, the median of a large number of DTW distances or average DTW distances acquired in one day. A series of representative evaluation values ​​is generated by the series processing unit 54 and input to the determination unit 55 as an initial series or a series for determination.

[0169] The transition of the log likelihood obtained in this embodiment is shown in the graph of Fig. 11. As shown in this graph, in this embodiment as well, as in the first embodiment, clear signs of a failure began to appear around December 16th.

[0170] As described above, in the robot failure sign detection device 5 of this embodiment, the evaluation value calculation unit 52 calculates, for each piece of current value time series data acquired by the current value time series data acquisition unit 51, either the DTW distance or the average DTW distance between the piece and a predetermined reference current value time series data as an evaluation value.

[0171] This makes it possible to easily grasp the change trend of the time-series data of the current value.

[0172] Next, a fourth embodiment will be described below. Fig. 12 is a block diagram showing an outline of the electrical configuration of a robot failure sign detection device 5 according to the fourth embodiment.

[0173] The robot failure sign detection device 5 of this embodiment corresponds to a modification of the first embodiment shown in FIG.

[0174] Similar to the first embodiment, the determination unit 55 of this embodiment includes a learning unit 57, a probability output unit 58, a determination model storage unit 59, and a conversion unit 61. The determination unit 55 further includes a probability storage unit 68 and a sum value output unit 69.

[0175] The learning unit 57 of this embodiment causes the hidden Markov model to learn the initial sequence, as in the first embodiment. The parameters of the obtained hidden Markov model are stored in the determination model storage unit 59.

[0176] As described above, in the first embodiment shown in Fig. 2, during the monitoring period (i.e., from day N+1 onwards), a series consisting of Q representative evaluation values ​​is input from the series processing unit 54 to the determination model of the determination unit 55 every day. On the other hand, in the present embodiment shown in Fig. 12, only one representative evaluation value obtained on that day is input from the storage unit 50 to the determination model of the determination unit 55 every day as determination data. In this embodiment, this one representative evaluation value corresponds to the determination data.

[0177] Generally, hidden Markov models are often used to evaluate sequential data. However, an ergodic hidden Markov model trained with an initial sequence can output values ​​that can be treated as probabilities even when a representative evaluation value is input as a single value rather than as a sequence in the evaluation phase.

[0178] The probability output unit 58 outputs the output result (i.e., probability) of the determination model when one representative evaluation value is input to the determination model. The probability output by the probability output unit 58 is stored in the probability storage unit 68.

[0179] The probability storage unit 68 can store the probabilities output by the determination model on or after the N+1 day for the most recent U days.

[0180] 13 shows the initial sequence and determination data in this embodiment when N=30 and U=15. On the 30th day, the learning unit 57 trains the hidden Markov model on the sequence of 30 representative evaluation values ​​shown in FIG. 13(a). This creates a determination model.

[0181] As shown in FIG. 13(b), on the 31st day, one representative evaluation value for the 31st day is input to the judgment model, and the probability output by the judgment model is stored. As shown in FIG. 13(c), on the 32nd day, one representative evaluation value for the 32nd day is input to the judgment model, and the probability output by the judgment model is stored. In FIG. 13, one circle represents one probability output by the judgment model. The number inside the circle indicates which day the probability corresponds to the representative evaluation value. If the circle is a dashed line, it indicates that the probability was stored in the past.

[0182] The determination unit 55 inputs one representative evaluation value obtained on that day into the determination model for each day and stores the probability output by the determination model. As a result, on the 45th day, 15 days' worth of probabilities are stored in the probability storage unit 68, as shown in Figure 13(d).

[0183] On the 45th day, the total product output unit 69 multiplies all of the probabilities for 15 days from the 31st day to the 45th day stored in the probability storage unit 68, as shown in FIG. 13(d). In this embodiment, the total product thus obtained corresponds to the specificity. The total product output by the total product output unit 69 is logarithmically transformed by the conversion unit 61 and output from the determination unit 55.

[0184] On the 46th day, the probability storage unit 68 deletes the oldest probability for the 31st day and stores the probability for the 46th day. The total product output unit 69 multiplies all of the probabilities for the 15 days from the 32nd day to the 46th day stored in the probability storage unit 68, as shown in FIG. 13( e). The total product thus obtained is logarithmically converted by the conversion unit 61 and output from the determination unit 55.

[0185] In this embodiment, output from the determination unit 55 essentially starts from the 45th day. Similar to the first and second embodiments, the configuration of this embodiment also makes it possible to properly detect signs of failure.

[0186] In this embodiment, the determination unit 55 outputs the sum of the logarithmic values, but the sum of the probabilities for the most recent 15 days may be output as the specificity. In this case, the logarithmic transformation may be omitted.

[0187] It is also possible to omit the probability storage unit 68. In this case, the model output values ​​from the 31st to the 45th days can be calculated all at once on the 45th day.

[0188] As described above, in this embodiment, after the initial stage of operation, the determination unit 55 inputs only one representative evaluation value as determination data to the determination model for each data collection unit period, and acquires the output of the determination model. The determination unit 55 acquires the specificity based on the sum or sum of the outputs of the determination model over the most recent multiple data collection unit periods.

[0189] This configuration also makes it possible to effectively detect signs of a malfunction in the robot 1.

[0190] Next, a fifth embodiment will be described. Fig. 14 is a block diagram showing an outline of the electrical configuration of a robot failure sign detection device 5 according to the fifth embodiment.

[0191] The robot failure sign detection device 5 of this embodiment further includes a normalization unit 71. The normalization unit 71 is used to normalize the representative evaluation value input to the hidden Markov model (determination model).

[0192] Now, let us explain normalization. Simply put, normalization is the process of converting the numerical values ​​contained in a certain data group so that the minimum value is 0, the maximum value is 1, and the values ​​in between are proportionally distributed between 0 and 1.

[0193] In this embodiment, the normalization unit 71 calculates the minimum and maximum values ​​of, for example, 30 representative evaluation values ​​during the initial operation period, and normalizes each representative evaluation value based on these values. The normalized values ​​are stored in the storage unit 50 and are used by the series processing unit 54 to create an initial series and a judgment series.

[0194] As another method of normalization, the data group may be transformed so that it has a normal distribution with a mean of 0 and a variance of 1.

[0195] As described above, current time-series data is acquired for the motors of each axis of the robot 1. In an actual robot example, the average current value of the wrist axis motor is about 3 A, while the average current value of the spindle motor is about 20 A. The magnitude of the current value naturally affects the representative evaluation value. As such, the representative evaluation value may differ greatly for each axis, and it is cumbersome to determine the threshold value that serves as the condition for issuing an alarm for each axis. Furthermore, the current value varies depending on the robot model, such as whether it is large or small, its payload capacity, and speed.

[0196] In this embodiment, these differences can be eliminated by normalization, allowing for unified judgment. The inventors of the present application have found that when the representative evaluation values ​​are normalized, the log likelihood falls within the range of -50 to 10 in the normal case, but when an abnormality occurs, the log likelihood often decreases to around -500. Therefore, setting a value of around -100 for the log likelihood as a unified threshold is considered to be generally appropriate for detecting signs of failure. In other words, individual anomaly judgment thresholds are no longer necessary.

[0197] As described above, in the robot failure sign detection device 5 of this embodiment, the representative evaluation value input to the determination model is normalized.

[0198] This allows a unified judgment to be made on the current value time series data obtained from various motors.

[0199] Next, a sixth embodiment will be described. Fig. 15 is a block diagram showing an outline of the electrical configuration of a robot failure sign detection device 5 according to the sixth embodiment.

[0200] 15, the robot failure sign detection device 5 of this embodiment includes two evaluation value calculation units 52, two representative evaluation value generation units 53, and two normalization units 71. In other respects, this embodiment is substantially similar to the first embodiment described above.

[0201] The two evaluation value calculation units 52 calculate the evaluation value of the current value time-series data using different functions. For example, one evaluation value calculation unit 52 may calculate the root mean square, and the other evaluation value calculation unit 52 may calculate the maximum value.

[0202] Accordingly, the hidden Markov model used in the determination unit 55 is configured to input a series of two-dimensional vectors consisting of two types of evaluation values ​​and calculate the probability, thereby making it possible to detect signs of failure by taking multiple evaluation values ​​into consideration in a composite manner.

[0203] A representative evaluation value representing each evaluation value is normalized by the normalization unit 71. This makes it possible to equalize the influence of each of the multiple evaluation values. Each normalized representative evaluation value may be multiplied by a factor determined for weighting. This makes it possible to determine a failure sign by placing more importance on the root mean square value than the maximum value, for example. The normalization unit 71 can also be omitted.

[0204] 15 shows a case where two types of evaluation values ​​are calculated for the current time-series data. However, the robot failure sign detection device 5 can also be configured to obtain three or more types of evaluation values. In this case, a series of three or more dimensional vectors is input to the hidden Markov model.

[0205] In the case of a normal distribution model, model f, i.e., equation (1), is a one-dimensional model, but in order to accommodate multiple dimensions, multiple models f, i.e., equation (1), can be prepared for each evaluation value and multiplied or added together, making it possible to accommodate multi-dimensional inputs in the same way as a hidden Markov model.

[0206] As described above, in the robot failure sign detection device 5 of this embodiment, the evaluation value calculation unit 52 calculates multiple types of evaluation values ​​using mutually different methods for each piece of current time series data acquired by the current time series data acquisition unit 51. The representative evaluation value generation unit 53 generates multiple types of representative evaluation values ​​for each data collection unit period. The judgment model is a model that can input a multi-dimensional series.

[0207] This makes it possible to detect signs of failure by taking multiple evaluation values ​​into consideration in a composite manner.

[0208] Furthermore, in the robot failure sign detection device 5 of this embodiment, the evaluation value calculation unit 52 calculates the root mean square and maximum value of the current value time series data. However, instead of one or both of the root mean square and maximum value, any one of the value range, the frequency analysis integrated value, the DTW distance, and the DTW distance average value may be calculated.

[0209] Furthermore, in the robot failure sign detection device 5 of this embodiment, the evaluation value calculation unit 52 calculates the root mean square and maximum value of the current value time series data. However, instead of both the root mean square and maximum value, two frequency analysis partial integrated values ​​may be calculated. It is also possible to configure the system to acquire three or more types of evaluation values. In this case, a series of three or more dimensional vectors is input to the hidden Markov model. Furthermore, as shown in FIG. 18, five partial integrated values ​​may be calculated.

[0210] This allows for a good evaluation of changes in the current value time series data.

[0211] Next, an example of predicting the time of failure of the robot 1 without using a judgment model will be described. The processing described in this example is performed by a robot maintenance support device 5a shown in FIG. 16. The robot maintenance support device 5a has the same basic configuration as the robot failure sign detection device 5. The robot maintenance support device 5a predicts future changes in the evaluation value based on the past trend of the evaluation value calculated based on the current value time series data, displays this on the display 80, and calculates the predicted failure date of the robot 1 based on the future changes in the evaluation value. This will be described in detail below with reference to FIGS. 16 and 17. FIG. 16 is an example of a trend management screen displayed on the display 80. FIG. 17 is a graph showing changes in the prediction line when the number of reference days is changed.

[0212] The trend management screen of FIG. 16 displays a graph display section 81, an evaluation value selection section 82, a diagnosis point selection section 83, and a predicted date display section 84.

[0213] The horizontal axis of the graph displayed on the graph display unit 81 represents time, and the vertical axis represents the evaluation value. The method for calculating the evaluation value is the same as in the above embodiment. The current time series data for calculating the evaluation value may be acquired while the robot 1 is performing actual work such as painting, or may be acquired while the robot 1 is performing a diagnostic operation. A diagnostic operation is an operation that is performed with the purpose of diagnosing the state of the robot 1. The reference line and the predicted line will be described later.

[0214] The evaluation value selection section 82 is a box for selecting the evaluation value of the graph displayed in the graph display section 81. The evaluation value selection section 82 displays the I2 monitor, DUTY, PTP, and frequency analysis integrated value, and the operator selects any one of them.

[0215] I2 is the root mean square of the current value. Root mean square is also called effective value or RMS. Root mean square of the current value represents the actual effect of the AC component. Therefore, I2 is a value that suppresses fine spike components. This allows for stable detection of increased loss in the reducer, i.e., a decrease in efficiency and a decrease in torque constant due to motor demagnetization. DUTY is the ratio of the motor's stall current to I2. PTP is an abbreviation for Peak to Peak and indicates the aforementioned "value range." In other words, PTP is the value obtained by subtracting the low peak current value from the high peak current value of the current waveform. PTP is commonly used because it is easy to calculate and can accurately estimate the state. The frequency analysis integrated value is the value described above. In the graphs of Figures 16 and 17, the frequency analysis integrated value is selected and displayed as the evaluation value.

[0216] The diagnosis point selection section 83 is a box for selecting a movable axis to be diagnosed from among the movable axes of the robot 1. In the diagnosis point selection section 83, a plurality of movable axes that the robot 1 has are displayed.

[0217] The predicted date display section 84 displays the predicted date of failure of the robot 1. The predicted date of failure is determined based on the following reference line and prediction line.

[0218] The reference line indicates the threshold value of the evaluation value. The threshold value is set, for example, to a value that is 120% of the effective value at the beginning of operation of the robot 1, or the initial value after the break-in of the robot 1. The value of 120% of the initial value after the break-in is an example, and a value other than 120% may be used. The threshold value may be empirically determined and registered. The threshold value is also changeable. Either initial value may be set from one piece of data, or an average value of multiple pieces of data may be used as the threshold value.

[0219] The prediction line is found by applying the least squares method to previously obtained evaluation values. The time when the prediction line intersects with the reference line is the predicted failure date. By looking at the prediction line, the operator can intuitively understand the state of robot 1. By looking at the predicted failure date, the operator can understand the specific date and time regarding the predicted failure of robot 1.

[0220] If the number of days until the predicted failure date approaches within a preset number of days (for example, 30 days), a warning may be displayed on the display 80. In this example, a prediction line is displayed for the selected evaluation value and movable axis, but the predicted failure date may also be found for combinations of evaluation value and movable axis that are not selected, and if the conditions are met, a warning may be displayed on the display 80.

[0221] Next, changing the display mode of a graph will be described with reference to FIG. 17. As shown in FIG. 17, one end of the horizontal axis of the graph display unit 81 is the drawing start date, and the other end of the horizontal axis is the drawing end date. The period from the drawing start date to the drawing end date corresponds to the display period of the evaluation value, forecast line, etc. In this example, the drawing start date and the drawing end date can be changed. For example, if the drawing start date is set to the first number of days before the current date and the drawing end date is set to the second number of days after the current date, these first and second number of days can be changed independently. This allows the operator to grasp the evaluation value range that he or she desires.

[0222] In this example, the number of reference days can also be changed. The number of reference days is the number of days that defines the range of evaluation values ​​used to calculate the forecast line. For example, if the number of reference days is 10 days, the forecast line is calculated based on the evaluation values ​​obtained from 10 days ago to the present. The number of reference days can be changed independently of the first and second numbers of days described above.

[0223] Below, we will explain what happens when the reference number of days is shortened to reference number of days a. When the reference number of days is shortened, the slope of the prediction line changes. Also, as a general trend, when the reference number of days is shortened, the slope of the prediction line becomes steeper. In Figure 17, the slope of prediction line a is also steeper than the slope of the prediction line. As a result of the change in the slope of the prediction line, the time when the prediction line intersects with the reference line changes, and therefore the predicted failure date changes. By having the function to change the reference number of days, the predicted failure date can be determined from various perspectives.

[0224] Furthermore, multiple partial integrated values ​​may be used instead of the frequency analysis integrated value. In order to accommodate multiple evaluation values, multiple trend management functions are provided. Figure 19 shows a trend management screen displayed in an example corresponding to a partial integrated value. Note that in Figure 19, descriptions of parts having the same functions as those in Figure 16 may be omitted.

[0225] The evaluation value selection section 82A in FIG. 19 displays the I2 monitor, DUTY, PTP, and partial integrated value. A pull-down menu for selecting a frequency band is displayed next to the partial integrated value. The frequency bands selectable using the evaluation value selection section 82A are 0-10, 10-20, 20-30, 30-40, 40-50, and ALL. ALL is the sum of all partial integrated values, and as a result, is displayed in the same way as the frequency analysis integrated value. The display method in FIG. 19 is one example and can be changed. For example, the frequency analysis integrated value may be added to the evaluation value selection section 82A, and ALL may be omitted from the menu for selecting a frequency band.

[0226] The reference line indicates the threshold value of the evaluation value. In the example shown in FIG. 16, it was explained that the threshold value is 120% of the initial value after the break-in of the robot 1. In the example shown in FIG. 19, the threshold value may be changed depending on the frequency band. For example, the threshold value when the frequency band is ALL or 0-10 may be 120% of the initial value, and the threshold value when the frequency band is 20-30, 30-40, or 40-50 may be 200% of the initial value. The prediction line is found by applying the least squares method to the evaluation values ​​obtained in the past for the reference period.

[0227] Furthermore, when calculating the number of days until the predicted failure date, the predicted failure date is calculated for any combination including not only the evaluation value and the movable axis, but also the frequency band. In the example shown in FIG. 19 , the predicted date display section 84A displays the overall predicted date in addition to the predicted dates described above. The overall predicted date may be the earliest combination of multiple predicted dates calculated based on multiple partial integrated values, or the earliest combination including other types of evaluation values ​​and movable axes. Furthermore, if the number of days until the overall predicted date is closer than a preset number of days and a warning is displayed, the screen may automatically switch to a screen displaying the combination of evaluation value, frequency band, and movable axis corresponding to the date listed in the overall predicted date. In this case, adding a message such as "Switch to the condition for which the warning was displayed" makes it easier for the operator to understand why the screen changed.

[0228] In this way, the robot maintenance support device 5a performs a process of acquiring time-series data of the drive current value for each movable axis of the robot 1 from the robot's operation for each data collection unit period. The robot maintenance support device 5a performs frequency analysis on each of the current time-series data to obtain a frequency spectrum, and then calculates a partial integrated value, which is a partial sum of the frequency spectrum, for multiple predetermined frequency bands. The robot maintenance support device 5a estimates the future change trend of each movable axis based on multiple partial integrated values ​​within a predetermined reference period. Furthermore, the robot maintenance support device 5a estimates the predicted lifespan of each movable axis based on the predicted time until the partial integrated value reaches a predetermined threshold. The robot maintenance support device 5a can also issue a warning based on the time (number of days) remaining until the predicted lifespan. The function of issuing a warning is not a required component and can be omitted.

[0229] The preferred embodiment of the present application has been described above, but the above configuration can be modified, for example, as follows.

[0230] When using an ergodic hidden Markov model, the initial series and the judgment series may have different numbers of representative evaluation values. For example, the initial series may have 30 representative evaluation values, while the judgment series may have 50 representative evaluation values. In this case, from the 31st to the 50th day, the series processing unit 54 only adds the latest representative evaluation value in the daily update process of the judgment series. From the 51st day onwards, the oldest representative evaluation value is deleted as the latest representative evaluation value is added. Input of the judgment series to the judgment model can begin on the 50th day. If it is acceptable for the judgment series to be shorter than usual, input of the judgment series to the judgment model can also begin on the 31st day.

[0231] The determination model based on the normal distribution of the second embodiment can also be combined with the fifth and sixth embodiments.

[0232] The DTW method of the third embodiment may be combined with the fourth, fifth, and sixth embodiments.

[0233] The graph of log-likelihood displayed on the display unit 63 may be displayed upside down. For example, upside down can be essentially achieved by multiplying the log-likelihood output by the conversion unit 61 by −1, inverting the sign, and outputting the resulting value to the display unit 63. This value increases as the similarity between the two sequences decreases, and is therefore easy to intuitively understand as a value representing the degree of anomaly.

[0234] It is possible to omit the conversion unit 61. That is, the probability output by the determination unit 55 can be compared with a threshold value or displayed on the display unit 63 without being logarithmically converted.

[0235] The robot failure sign detection device 5 can also be realized by the same hardware as the controller 90.

[0236] The functions of the elements disclosed herein can be performed using circuits or processing circuitry, including general-purpose processors, special-purpose processors, integrated circuits, ASICs (Application Specific Integrated Circuits), conventional circuits, and / or combinations thereof, configured or programmed to perform the disclosed functions. A processor is considered a processing circuit or circuitry because it includes transistors and other circuitry. In this disclosure, a circuit, unit, or means is hardware that performs the recited functions or hardware that is programmed to perform the recited functions. The hardware may be hardware disclosed herein or other known hardware that is programmed or configured to perform the recited functions. Where the hardware is a processor, which is considered a type of circuit, the circuit, means, or unit is a combination of hardware and software, and the software is used to configure the hardware and / or processor.

Claims

1. A robot failure sign detection device for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition unit that performs a process of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation unit that calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit; a representative evaluation value generation unit that generates, from the evaluation values ​​obtained by the evaluation value calculation unit, a representative evaluation value that represents the evaluation values, for each data collection unit period; a series processing unit that generates a series of the representative evaluation values; a determination unit that, in an initial stage of operation that indicates an initial stage after installation of the robot, creates a determination model based on an initial series that is a series generated by the series processing unit, and, after the initial stage of operation, inputs a determination series that includes data based on robot operations after the initial stage of operation and is made up of a plurality of the representative evaluation values ​​into the determination model as determination data for each data collection unit period, and obtains the specificity of the determination data based on the output of the determination model; Equipped with In order to generate the judgment sequence, the sequence processing unit initializes the judgment sequence with the initial sequence, and then performs processing to update the judgment sequence every time the representative evaluation value is generated; The robot failure sign detection device, wherein the update process includes a process of adding the representative evaluation value for the data collection unit period to the judgment series before the update.

2. A robot failure sign detection device for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition unit that performs a process of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation unit that calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit; a representative evaluation value generation unit that generates, from the evaluation values ​​obtained by the evaluation value calculation unit, a representative evaluation value that represents the evaluation values, for each data collection unit period; a series processing unit that generates a series of the representative evaluation values; a determination unit that, in an initial stage of operation that indicates an initial stage after installation of the robot, creates a determination model based on an initial series that is a series generated by the series processing unit, and, after the initial stage of operation, inputs determination data that includes data based on robot operations after the initial stage of operation into the created determination model, and obtains the specificity of the determination data; Equipped with After the initial stage of operation, the judgment unit inputs only one of the representative evaluation values ​​as the judgment data into the judgment model for each data collection unit period to obtain an output of the judgment model, and obtains the specificity based on the sum or sum of the outputs of the judgment model over a plurality of most recent data collection unit periods.

3. A robot failure sign detection device for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition unit that performs a process of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation unit that calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit; a representative evaluation value generation unit that generates, from the evaluation values ​​obtained by the evaluation value calculation unit, a representative evaluation value that represents the evaluation values, for each data collection unit period; a series processing unit that generates a series of the representative evaluation values; a determination unit that, in an initial stage of operation that indicates an initial stage after installation of the robot, creates a determination model based on an initial series that is a series generated by the series processing unit, and, after the initial stage of operation, inputs determination data that includes data based on robot operations after the initial stage of operation into the created determination model, and obtains the specificity of the determination data; Equipped with The robot failure sign detection device, wherein the judgment model is a hidden Markov model that has been trained on the initial sequence.

4. 4. The robot failure sign detection device according to claim 3, the evaluation value calculation unit calculates a plurality of types of evaluation values ​​using mutually different methods for each of the behavior time-series data acquired by the behavior time-series data acquisition unit, the representative evaluation value generation unit generates a plurality of types of the representative evaluation values ​​for each data collection unit period; The robot failure sign detection device, wherein the judgment model is a model that can input a multi-dimensional series as the initial series.

5. A robot failure sign detection device for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition unit that performs a process of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation unit that calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit; a representative evaluation value generation unit that generates, from the evaluation values ​​obtained by the evaluation value calculation unit, a representative evaluation value that represents the evaluation values, for each data collection unit period; a series processing unit that generates a series of the representative evaluation values; a determination unit that, in an initial stage of operation that indicates an initial stage after installation of the robot, creates a determination model based on an initial series that is a series generated by the series processing unit, and, after the initial stage of operation, inputs determination data that includes data based on robot operations after the initial stage of operation into the created determination model, and obtains the specificity of the determination data; Equipped with The judgment model is a robot failure sign detection device that is formed by multiplying or summing Q times a normal distribution based on the standard deviation and mean value obtained from the N representative evaluation values ​​that make up the initial series.

6. 6. The robot failure sign detection device according to claim 5, the evaluation value calculation unit calculates a plurality of types of evaluation values ​​using mutually different methods for each of the behavior time-series data acquired by the behavior time-series data acquisition unit, the representative evaluation value generation unit generates a plurality of types of the representative evaluation values ​​for each data collection unit period; The robot failure sign detection device, wherein the judgment model is a model that can input a multi-dimensional series as the initial series.

7. 7. The robot failure sign detection device according to claim 1, The evaluation value is any one of the root mean square, maximum value, value range, and frequency analysis integrated value of the behavior time series data.

8. A robot failure sign detection device for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition unit that performs a process of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation unit that calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit; a representative evaluation value generation unit that generates, from the evaluation values ​​obtained by the evaluation value calculation unit, a representative evaluation value that represents the evaluation values, for each data collection unit period; a series processing unit that generates a series of the representative evaluation values; a determination unit that, in an initial stage of operation that indicates an initial stage after installation of the robot, creates a determination model based on an initial series that is a series generated by the series processing unit, and, after the initial stage of operation, inputs determination data that includes data based on robot operations after the initial stage of operation into the created determination model, and obtains the specificity of the determination data; Equipped with The evaluation value calculation unit calculates, as the evaluation value, either the DTW distance or the average DTW distance between each of the behavior time series data acquired by the behavior time series data acquisition unit and a predetermined reference behavior time series data.

9. 7. The robot failure sign detection device according to claim 4 or 6, A robot failure sign detection device, wherein the multiple types of evaluation values ​​include any of the root mean square, maximum value, value range, frequency analysis integrated value, DTW distance, and DTW distance average value of the behavior time series data.

10. A robot failure sign detection device for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition unit that performs a process of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation unit that calculates an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition unit; a representative evaluation value generation unit that generates, from the evaluation values ​​obtained by the evaluation value calculation unit, a representative evaluation value that represents the evaluation values, for each data collection unit period; a series processing unit that generates a series of the representative evaluation values; a determination unit that, in an initial stage of operation that indicates an initial stage after installation of the robot, creates a determination model based on an initial series that is a series generated by the series processing unit, and, after the initial stage of operation, inputs determination data that includes data based on robot operations after the initial stage of operation into the created determination model, and obtains the specificity of the determination data; Equipped with the evaluation value calculation unit performs frequency analysis on each of the behavior time-series data acquired by the behavior time-series data acquisition unit to obtain a frequency spectrum, and calculates a plurality of partial sums of the frequency spectrum as the evaluation values ​​for a plurality of predetermined frequency bands; the representative evaluation value generation unit generates a plurality of the representative evaluation values ​​for each data collection unit period; The robot failure sign detection device, wherein the judgment model is a model that can input a multi-dimensional series as the initial series.

11. 10. The robot failure sign detection device according to claim 9, At least one of the representative evaluation values ​​input to the judgment model is normalized.

12. A robot failure sign detection method for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition step of acquiring behavior time-series data relating to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation step of calculating an evaluation value for the behavior time-series data acquired in the behavior time-series data acquisition step; a representative evaluation value generating step of generating, from the evaluation values ​​obtained in the evaluation value calculating step, a representative evaluation value that represents the evaluation values ​​for each data collection unit period; a sequence processing step for generating a sequence of the representative evaluation values; a model creation step of creating a judgment model based on the initial sequence generated in the sequence processing step at an initial stage after installation of the robot; a determination step of, after the initial stage of operation, inputting, for each data collection unit period, a determination series including data based on robot operations after the initial stage of operation and consisting of a plurality of the representative evaluation values ​​as determination data into the determination model, and acquiring a specificity of the determination data based on an output of the determination model; In order to generate the judgment sequence, the judgment sequence is initialized with the initial sequence, and then an update process of the judgment sequence is performed every time the representative evaluation value is generated; The robot failure sign detection method, wherein the update process includes a process of adding the representative evaluation value for the data collection unit period to the judgment series before updating.

13. A robot failure sign detection method for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition step of acquiring behavior time-series data relating to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation step of calculating an evaluation value for the behavior time-series data acquired in the behavior time-series data acquisition step; a representative evaluation value generating step of generating, from the evaluation values ​​obtained in the evaluation value calculating step, a representative evaluation value that represents the evaluation values ​​for each data collection unit period; a sequence processing step for generating a sequence of the representative evaluation values; a model creation step of creating a judgment model based on the initial sequence generated in the sequence processing step at an initial stage after installation of the robot; a determination step of inputting, after the initial stage of operation, determination data including data based on robot operations after the initial stage of operation into the created determination model, and acquiring a specificity of the determination data; In the judgment step, for each data collection unit period, only one of the representative evaluation values ​​is input to the judgment model as the judgment data to obtain an output of the judgment model, and the specificity is obtained based on the sum or sum of the outputs of the judgment model over the most recent multiple of the data collection unit periods.

14. A robot failure sign detection program for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition step of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation step of calculating an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition step; a representative evaluation value generating step of generating, from the evaluation values ​​obtained in the evaluation value calculating step, a representative evaluation value that represents the evaluation values ​​for each data collection unit period; a series processing step for generating a series of the representative evaluation values; a model creation step of creating a judgment model based on an initial sequence, which is a sequence generated in the sequence processing step, in an initial stage of operation after installation of the robot; a determination step of inputting, for each data collection unit period after the initial stage of operation, a determination series including data based on robot operations after the initial stage of operation and consisting of a plurality of the representative evaluation values ​​as determination data into the determination model, and acquiring a specificity of the determination data based on an output of the determination model; an updating step of initializing the judgment sequence with the initial sequence to generate the judgment sequence, and then updating the judgment sequence every time the representative evaluation value is generated; It is a program that causes a computer to execute The updating step includes adding the representative evaluation value of the data collection unit period to the judgment series before updating.

15. A robot failure sign detection program for detecting a failure sign of a robot, comprising: a behavior time-series data acquisition step of acquiring behavior time-series data related to the motors of the joints of the robot from the robot's motion for each data collection unit period; an evaluation value calculation step of calculating an evaluation value for the behavior time-series data acquired by the behavior time-series data acquisition step; a representative evaluation value generating step of generating, from the evaluation values ​​obtained in the evaluation value calculating step, a representative evaluation value that represents the evaluation values ​​for each data collection unit period; a series processing step for generating a series of the representative evaluation values; a model creation step of creating a judgment model based on an initial sequence, which is a sequence generated in the sequence processing step, in an initial stage of operation after installation of the robot; a determination step of inputting, after the initial stage of operation, determination data including data based on robot operations after the initial stage of operation into the created determination model, and acquiring a specificity of the determination data; It is a program that causes a computer to execute a robot failure sign detection program for detecting a failure sign in a robot, the robot failure sign detection program inputting only one of the representative evaluation values ​​as the judgment data into the judgment model for each data collection unit period to obtain an output of the judgment model, and obtaining the specificity based on the sum or sum of the outputs of the judgment model over the most recent multiple of the data collection unit periods in the judgment step;

Images