WO2019057356A1

WO2019057356A1 - Anomaly detection in an industrial robot

Info

Publication number: WO2019057356A1
Application number: PCT/EP2018/066429
Authority: WO
Inventors: Fan Dai; Debora CLEVER; Boris FIEDLER; Benjamin Klöpper; Jan-Christoph SCHLAKE; Marcel Dix; Subanatarajan Subbiah
Original assignee: Abb Schweiz Ag
Priority date: 2017-09-21
Filing date: 2018-06-20
Publication date: 2019-03-28

Abstract

A method for anomaly detection in an industrial robot, comprising the steps of a) while executing a predetermined movement, recording (S11) a time series of at least one first operating parameter of the robot; b) applying a machine learning model to the time series obtained in step a), and c) use the machine learning model to score the time series obtained in step a), deciding whether it is an anomaly or not.

Description

Anomaly detection in an industrial robot

The present invention relates to a method and appa- ratus for detecting an anomaly in an industrial ro^¬ bot .

An anomaly is the deviation of an observed operat^¬ ing parameter or a time series of an observed oper- ating parameter from a desired behaviour.

An example for such an anomaly is for example an unexpectedly increased friction in a joint of the robot or any other symptom that might forebode a defect.

Such anomalies should be detected as early as pos^¬ sible, in order to minimize consequential damage. Further, if a need for future repair can be fore- cast, preparations for the repair can be taken while the robot is still operating, so that down^¬ time can be minimized.

If for example an unexpectedly increase in friction is the anomaly to be detected, friction in a given joint of the robot can be estimated when the orien^¬ tation of the axis of the joint and the torque act^¬ ing on the joint are known. The former can be cal^¬ culated based on the angles of all proximal joints located between said given joint and a base of the robot, the latter can be calculated based on the weights of links on the distal side of said given link and associated joint angles.

However, these calculations require significant computing power which adds to the overall cost of the robot .

In a more generic way, there is also the aspect of storage cost and the aspect of possible transmis- sion of the recoded data of the observed parame^¬ ter. If a lot of data is recorded, and it needs to be stored in the robot, this increases the cost of the robot . An object of the invention is, therefore, to pro^¬ vide a method for anomaly detection in an industrial robot which requires little processing power and reduced data recording. The object is achieved by a method for anomaly de^¬ tection in an industrial robot, comprising the steps of

a) while executing a predetermined movement, re^¬ cording a time series of at least one first operat- ing parameter of the robot;

b) applying a machine learning model to the time series obtained in step a) , and

c) use the machine learning model to score the time series obtained in step a) , deciding whether it is an anomaly or not.

Applying the machine learning model approach also takes the technical approach to solve the problem of distinguishing between a deviation of an operat- ing parameter which caused by a "normal" variation in the movement and an "un-normal" variation due to a technical problem. Key ability of a machine learning model is to generalize. Some examples of possible machine learning model^¬ ling methods that are known and can be applied here comprise instance-based (K-nearest neighbor or Lo- cal-outlier-factor ), or model-generation (One-class svm, Stacked-Auto-Encoder ) .

Machine learning can be split into two phases: A first phase, which is a training phase, and a sec- ond phase, which is a scoring phase.

The training requires to obtain a training data set containing historic data. The second step takes one sample, for example pro^¬ duced by the most recent robot movement, and uses the model to score the sample, i.e. to decide whether it is an anomaly or not. So according to the invention, the decision if the time-series from step a) is anomalous or not, is made by a machine learning model. This machine learning model has been previously trained on his^¬ toric data from the same robot or similar robots with similar applications respectively movement program .

In an advantageous embodiment of the invention, step b) comprises determining a deviation between the time series obtained in step a) and a model time series; and step c) comprises deciding that there is an anomaly if the deviation exceeds a giv^¬ en threshold. The method according to the invention is based on the assumption that the movement of a robot tends to be highly repetitive. Therefore, it isn't always necessary to calculate an operating parameter in which an anomaly might manifest itself; rather, if a model time series of the operating parameter has been obtained once for the repetitive movement, it is sufficient, whenever the movement is repeated, to record a time series of the operating parameter and to check for possible deviations from the model time series. The model time series may be calculated as de^¬ scribed above, preferably it is based on measure^¬ ment instead. To this effect, step a) can be exe^¬ cuted at least twice, and the model time series is derived from the time series recorded in the first execution of step a) .

The influence of measurement errors may be reduced if the model time series is derived from a plurali^¬ ty of recorded time series, e.g. by averaging.

If the movement of a robot is programmed as a se^¬ ries of coordinate or speed vector instructions, a controller which receives these instructions cannot tell where the movement begins to repeat itself, and therefore can only monitor the series of in^¬ structions it receives for repetitive patterns. In the meantime, first operating parameters must be recorded, in order to have data available from which a time series can be extracted in case the movement proves repetitive. This again requires a lot of processing power. On the other hand, if the instructions received specify not merely a coordi^¬ nate or a speed, but a task which the controller translates into speed or position instructions for execution by the robot, it is highly likely that a movement embodied in the task instruction will be repetitive, so that the need for processing power can be reduced if recording the time series is be^¬ gun in synchronism with the beginning of the task.

Further, the definition of a task may specify a first operating parameter which should be recorded when the task is being executed. This can also sim^¬ plify processing, since recording can be limited to those operating parameters whose variation is in^¬ deed correlated with the execution of the task.

The first operating parameter shouldn't be subject to closed-loop control, because in that case, even if there is an anomaly, the controller will vary some control parameter until the first operating parameter fits its target value, rather than allow the first operating parameter to deviate. Rather, if a second operating parameter is closed-loop con^¬ trolled while executing the predetermined task, and a relation between the first and second operating parameters is variable by an anomaly, the existence of an anomaly should be clearly discernible in the values of the first operating parameter.

For instance, one of the operating parameters may be related to the speed or the position of a link of the robot, and the other operating parameter is related to the power or the speed of a motor driv^¬ ing the link. In that case, if e.g. the speed of the link is closed-loop controlled, the power of the motor which the controller will adapt in the attempt to achieve the desired speed of the link will vary strongly depending on the friction of the joint, and thus betray any anomaly in the coupling between the motor and the link driven by it.

If above-defined steps a) to c) are carried out for a plurality of first operating parameters, a warn- ing should be output in step c) which is indicative of the first operating parameter that gave rise to the decision, so that the nature of the anomaly can be judged, and appropriate measures can be taken for removing it.

The object is also achieved by an apparatus for anomaly detection in an industrial robot, the appa^¬ ratus comprising

detecting means for detecting a time series of at least one first operating parameter of the ro^¬ bot ;

storage means for storing a machine learning model and a training data set containing historic data,

decision means for using the machine learning model to score the detected time series and decid^¬ ing whether it is an anomaly or not.

In an advantageous embodiment, the apparatus com^¬ prises

storage means in which a model time series of the at least one first operating parameter is stored; and

decision means for comparing the detected time series and the model time series and for deciding that there is an anomaly if the deviation exceeds a given threshold.

The object is also achieved by a computer program product comprising program code means which enable a computer to perform the method as defined above.

Further features and advantages of the invention will become apparent from the following description of embodiments thereof, referring to the appended drawings . Fig. 1 schematically illustrates a robot system according to the invention.

Fig. 2 is a diagram illustrating operating parameters of the robot system as a function of time ;

Fig. 3 is another diagram illustrating operating parameters of the robot system as a func^¬ tion of time; and

Fig. 4 is a flowchart of an operating method of the system of Fig. 1.

Fig. 1 illustrates a robot system comprising a ro^¬ bot arm 1 and its associated controller 2. The ro^¬ bot arm 1 comprises a support 3, an end effector 4 an arbitrary number of links 5 and joints 6 which connect to the links 5 to each other, to the sup^¬ port 3 or to the end effector 4 and have one or two degrees of rotational freedom. As usual in the art, motors for driving rotation of the links 5 and the end effector 4 about axes 7, 8 are hidden inside the links 5, the joints 6 or the support 3. The joints 6 further comprise rotary encoders or other appropriate sensors associated to each axis 7, 8 which provide the controller 2 with data on the orientation of each link 5 and on the distance between grippers 9 of end effector 4, and torque sen^¬ sors which are sensitive to torque in the direction of axis 7 and 8, respectively.

In the present example, it will be assumed that the robot's job is to seize a screw 10 from a table top and screw it into a bore of a workpiece 11. The po^¬ sition and the orientation of the screw 10 is not defined in advance, so the controller 2 uses a cam- era 12 to identify the screw 10 and to place the end effector 4 so that an axis of rotation 13 of the end effector 4 coincides with a longitudinal axis 14 of the screw 10 and the grippers 9 of end effector 4 are spaced under a predetermined opening angle Θ0 which is wide enough so that the head of screw 10 fits loosely between the grippers 9. Since the position and orientation may vary from one screw 10 to the next, the operating parameters of the motors of the links 5 which place the end ef^¬ fector 4 next to screw 10 on the table top will al^¬ so vary. However, when the end effector 4 has been placed appropriately, the torque required to close the grippers 9 and seize the screw 10 depends but little on the orientation of the end effector, since the grippers 9 carry no weight but their own. Therefore, the controller 2 may then control a gripper drive motor to approach the grippers 9 to each other at a predetermined speed, and in the meantime, record a time series of the input elec^¬ tric current required by the motor for attaining the predetermined speed, or apply a predetermined current to the motor and monitor a time series of the resulting speed of the grippers 9.

A time series of data thus obtained is subsequently compared to a previously established model time se^¬ ries. Typically, the model time series may have a shape similar to curve CI of Fig. 2: while the grippers 9 are approaching the head of the screw, the torque required for approaching the grippers 9 is low and varies little, since only internal fric^¬ tion of the end effector 4 has to be overcome. When the angle between the grippers 9 has been reduced to 01 and the grippers 9 come into contact with the screw head, the torque rises abruptly, and dis^¬ placement of the gripper 9 ends. A currently recorded time series may differ from the model time series in various ways and for vari^¬ ous reasons. These reasons may be external or in- ternal to the robot arm 1. External reasons may be related to the objects to be seized. E.g. if the screw 10 is seized not from a free table top but from a pile of screws, not only the targeted screw head, but also part of another screw may be between the grippers 9. So there may be a premature rise of the torque at some angle Θ2 when this other part gets squeezed, most likely followed by a breakdown of the torque when both screws escape from between the grippers 9, as shown by curve C2. If the grip- pers 9 continue to approach each other, the torque will rise again when the angle Θ between them becomes zero, i.e. if the grippers 9 touch each oth^¬ er . Internal problems such as bearing damage, insuffi^¬ cient lubrication or breakage may cause the torque to vary unpredictably, to be constantly higher than in the model, as exemplified by curve C3, or to be constantly close to zero, respectively. So by com- paring the shape of the curve derived from the time series obtained during the closing movement of the grippers 9 to that of the model, the existence and, eventually, even the cause of an irregularity can be detected.

Similarly, and as shown in Fig. 3, while screw 10 is being screwed into a threaded bore of workpiece 11 by rotating end effector 4 around axis 13, the torque exercised on the screw 10 is governed by friction between the screw 10 and the bore and should not exceed a predetermined threshold before, at a rotation angle φΐ, the head of screw 10 abuts against workpiece 11, as shown in curve CI'. If there is a premature rise of the torque at e.g. ro^¬ tation angle φ2, it might be due to an external cause such as defective threads of the screw 10 or the workpiece 11, bad alignment of the axes or the like. Internal problems can be excessive friction in the end effector 4 (curve C3 ' ) , as in case of Fig. 2. If the torque begins to increase at φΐ but, instead of rising to the same level as in curve CI', breaks down as shown in curve C'4, the screw 10 might not be held tight enough, and the grippers 9 might be slipping while rotating around its head.

Fig. 4 is a flowchart of operations of controller 2 in the process of seizing a screw 10 and screwing it into workpiece 11.

Based on data from camera 12, controller 2 chooses a screw to be seized and determines its position and orientation (SI) . It places the grippers 9 at an angle Θ0 in which the space between them is comfortably large enough for the head of the screw 10 to fit in, aligns axes 13, 14 in parallel with each other and places the end effector 4 over screw 10 so that the two axes coincide (S2) .

In step S3, the controller begins the task of actu^¬ ally seizing the screw 10. This can be done by branching to a screw seizing subroutine. Branching to the subroutine triggers collection of values of a predetermined operating parameter of the end ef^¬ fector 4. For example, if the controller 2 controls the grippers 9 to approach each other at a prede^¬ termined angular velocity d0/dt, then the recorded operating parameter can be the electric current I(t) that must be input into the motor driving the grippers 9 to achieve this angular velocity d0/dt. While the grippers 9 are approaching each other, the controller 2 repeatedly records this electric current at predetermined intervals of time t or an- gle Θ (S4), in order to obtain a time series of the current intensity. The recording is stopped (S5) when the current I (ti) has risen to a value Imax corresponding to a predetermined gripping torque Tmax ( cf . Fig . 2 ) .

Steps SI to S5 can be repeated for a plurality of screws, and the resulting time series can be aver^¬ aged in step S6. At this time the seizing process should still be supervised by a human in order to discard time se^¬ ries obtained in failed seizing attempts.

The result of step S4 or, eventually, S6 is a time series which in step S7 is stored as a model I in a solid state memory 16 associated to controller 2.

After the model has been established, the steps of choosing a screw and placing the end effector and recording currents are carried out as before (S8- Sll) . For each current value recorded in step Sll, a deviation Δ_± from the model I is calculated in step S12, for example

i

While the accumulated deviation ^Δ_¾ is found to be k=l

below a predetermined threshold in step S13, the movement of the grippers 9 and the collection of current data I± is continued until the screw has been seized. When that happens, the accumulated de^¬ viation is reset to zero (S14), and the process is interrupted until another screw is to be seized (corresponding to the dashed line in Fig. 4) . If the threshold is exceeded, it is decided in step S15 that there is an anomaly, and a warning that something is wrong with the grippers 9 is output at a user interface 15. The warning may comprise a guess as to the reason for the anomaly, which is derived from the shape of the curve derived from the time series, as explained referring to Fig. 2.

The task of screwing the screw 10 into workpiece 11 is monitored by a similar method. The controller 2 places the end effector 4 so that the tip of the screw 10 held by it is located at the opening of the bore in workpiece 11, and the axes 13, 14 are aligned with the bore (S2), and starts to rotate the screw 10 and to record current intensities of the motor that drives the rotation of the end ef- fector4. When doing so for the first time (or times), motor currents are recorded at predeter^¬ mined intervals of time or rotation angle φ in or- der to derive a model time series (S4-S7) .

Once the model time series is established, subse^¬ quently detected time series are compared to it. Deviations detected in step S12 can be due to the screw, to the bore or to internal defects of the robot arm, in particular of the motor which rotates the end effector by axis 13, so that if these devi^¬ ations exceed a threshold, an anomaly is diagnosed in step S15.

It should be kept in mind that in the above con^¬ text, a time series must not be construed as a set of data collected at regular time intervals. In the above examples, if the grippers 9 or the end effec- tor 4 is controlled to rotate at a constant angular velocity, data which are collected at regular in^¬ tervals of the rotation angle Θ or φ happen to be regularly spaced in time, too. However, no such regular spacing is required; it should only be en^¬ sured that the pairs of data from the model time series and the current time series are associated to a same value of the rotation angle Θ or φ, re^¬ spectively .

It should also be kept in mind that the choice of angular velocity and motor current intensity as in- terrelated operating parameters of the robot in the above examples is quite arbitrary, and that, de^¬ pending on the task to be executed, other parame^¬ ters might be just as appropriate or even more so. For instance, the grippers 9 might be equipped with pressure sensors, e.g. piezo sensors, for measuring directly the pressure exerted on a seized object; in that case this pressure might be a suitable op^¬ erating parameter for the task of seizing a screw or other workpiece.

Reference numerals

1 robot arm

2 controller

3 support

4 end effector

5 link

6 joint

7 axis

8 axis

9 gripper

10 screw

11 workpiece

12 camera

13 axis

14 axis

15 user interface

16 memory

Claims

1. A method for anomaly detection in an industrial robot, comprising the steps of

a) while executing a predetermined movement, recording (Sll) a time series of at least one first operating parameter of the robot;

2. The method of claim 1, wherein step b) comprises determining (S12) a deviation ( Δ,∑) between the time series obtained in step a) and a model time series; and step c) comprises de^¬ ciding (S15) that there is an anomaly if the deviation ( Δ,∑) exceeds (S13) a given threshold .

3. The method of claim 2, wherein step a) is exe^¬ cuted at least twice (S4, Sll), and the model time series is derived (S4, S6) from the time series recorded in the first execution (S4) of step a) .

4. The method of claim 3, wherein the model time series is derived (S6) from a plurality of recorded time series.

5. The method of any of the preceding claims, wherein the movement is predefined in a task, and recording (S4, Sll) the time series is be- gun (S3, S10) in synchronism with the beginning of the task .

The method of any of the preceding claims, wherein a second operating parameter is closed-loop controlled while executing the predetermined task, and a relation between the first and second operating parameters is vari^¬ able by an anomaly.

The method of claim 6, wherein one of the op^¬ erating parameters is related to the speed or the position of a link of the robot, and the other operating parameter is related to the power or the speed of a motor driving the link .

The method of any of the preceding claims, wherein steps a) to c) are carried out for a plurality of first operating parameters, and in step c) a warning is output (S15) which is indicative of the first operating parameter that gave rise to the decision.

An apparatus for anomaly detection in an industrial robot, comprising

detecting means for detecting a time series of at least one first operating parameter of the robot;

decision means for using the machine learning model to score the detected time se^¬ ries and deciding whether it is an anomaly or not .

10. An apparatus according to claim 9, comprising storage means (16) in which a model time series of the at least one first operating pa^¬ rameter is stored;

- decision means for comparing the detected time series and the model time series and for deciding that there is an anomaly if the devi^¬ ation exceeds a given threshold.

11. A computer program product comprising program code means which enable a computer to perform the method according to any of claims 1 to 8.