DE102019214009B3 - Analysis of sensor measurement data from a robot or manipulator - Google Patents

Abstract

According to a method according to the invention for analyzing sensor measurement data from a robot or manipulator, the following steps are carried out:
- Receiving or determining a data record of sensor measurement data that represent a time course of a measured variable;
- Identifying at least one at least approximately repeating pattern in the time course of the measured variable in the data set;
Subdividing the data set into several sections, each section corresponding to a time section which has at least approximately the duration of the at least approximately repeating pattern; and
Determination of the sections in which the at least approximately repeating pattern or a pattern deviating therefrom within a tolerance range occurs.

Description

The present invention relates to the analysis of sensor measurement data from a robot or manipulator.

It is known from the prior art to collect sensor measurement data from a robot or manipulator that represent a time profile of a measured variable. Such a measured variable can be, for example, a position of a movable member of a manipulator or robot, in particular a single-axis or multi-axis robot. The measured variable could also be a first or higher time derivative of such a position (speed, acceleration ...). Further examples of suitable measured variables are forces that act on or on a member of the actuator or robot, in particular between two members of a robot, or torques on or between such members. Such measured variables are only mentioned by way of example. In principle, the present invention can also be used in relation to other measured variables that can be recorded in or on a robot or manipulator. In the following, only the term “measured variable” is used for the sake of simplicity, without it being pointed out every time that it is a measured variable of a robot or manipulator or a measured variable that can be measured in or on a robot or manipulator .

The DE 11 2016 001 831 B4 relates to a method and a device for controlling / regulating a robot joint driven by an electric motor, the robot joint having a current sensor with first sensor electronics for detecting a first operating current of the electric motor, a first position sensor for detecting a drive position of a drive train of the robot joint, a has a second position sensor for detecting an output position of an output train of the robot joint, and a first torque sensor for detecting a torque in the output train, the electric motor being controlled / regulated on the basis of a predetermined setpoint manipulated variable.

The DE 10 2015 211 348 A1 relates to a handling device and a method for gripping, detecting, positioning, handling, shifting, rotating, regrouping and / or manipulating articles, bundles or piece goods transported individually, in pairs or in groups on a horizontal conveyor, the handling device being a movably suspended and in a movement space above the horizontal conveyor device, which can be moved in a controlled manner with one or more means for gripping the articles, groupings, bundles or piece goods.

The DE 10 2014 103 240 A1 relates to a method for setting up and / or monitoring operating parameters of a workpiece processing machine, which has a tool holder and means for moving a workpiece and the tool holder relative to one another, in any case along a first axis, with the tool holder equipped with a tool in the machining operation and when the tool engages a Workpiece values for at least one of the following measured variables occurring during interaction between the tool and the workpiece on the tool and transmitted to the tool holder are recorded on the tool holder and recorded for the processing sequence: a. an axial force acting in a direction parallel to the first axis, b. a torque applied in relation to the first axis or an axis lying parallel thereto, c. Bending moments or bending moment components according to direction and amount, the determined values for the at least one measured variable being used to set up the operating parameters and / or to adjust the processing process with regard to a longer service life of the tool used while at the same time falling below a maximum processing time The reproducibility of the same and / or a tool wear and / or a machine error of the workpiece processing machine to be monitored.

The DE 10 2008 054 312 A1 relates to a method for the reliable detection of a kinematic variable of a manipulator, comprising the steps of: detecting a kinematic variable of the manipulator at different points in time by a detection device, determining a change in the detected kinematic variable over time, checking whether the change over time is in a predetermined plausible range is, determining a plausibility of the detected kinematic variable on the basis of this check and outputting a safe kinematic variable on the basis of the detected kinematic variable if the change over time is in the plausible range.

The EP 2 422 935 B1 relates to a robot comprising an arm and a strain sensor comprising a piezoelectric body whose natural frequency is higher than the natural frequency of a structural material forming the arm.

The US 2014/0201571 A1 relates to a system for condition monitoring and fault diagnosis comprising a data acquisition function that records the time courses of selected variables for one or more of the components, a Preprocessing function that calculates certain features of the time courses, an analysis function for evaluating the features, generating one or more hypotheses of a state of the one or more components and an argumentation function for determining the state of the one or more components from the one or more hypotheses.

The inventors have recognized that conclusions can be drawn from the measured variables collected, for example in the context of controlling or monitoring a robot or manipulator or diagnosing problems that may occur with robots or manipulators. Relatively simple techniques are already known from the prior art, in which measured values are used for the control, monitoring or diagnosis of robots or manipulators. As a simple example, it should be mentioned here that a torque measured on an axis of a robot can be compared with a maximum permissible torque. If the measured torque exceeds the maximum permissible torque, an alarm can be triggered or an error message can be output, for example.

The object of the present invention is to provide an alternative, in particular improved method for analyzing sensor measurement data from a robot or manipulator.

This object is achieved by a method with the features of claim 1. Claim 13 provides a device (control, monitoring, analysis or diagnosis device) under protection, which is set up to carry out a method described here. Claims 14 and 15 provide a computer program or a computer program product for performing a method described here under protection. The subclaims relate to advantageous developments.

• - Empfangen oder Ermitteln eines Datensatzes von Sensormessdaten, die einen zeitlichen Verlauf einer Messgröße darstellen;
• - Identifizieren wenigstens eines sich wenigstens näherungsweise wiederholenden Musters im zeitlichen Verlauf der Messgröße in dem Datensatz;
• - Unterteilen des Datensatzes in mehrere Abschnitte, wobei jeder Abschnitt einem Zeitabschnitt entspricht, der wenigstens näherungsweise die Dauer des sich wenigstens näherungsweise wiederholenden Musters hat; und
• - Ermitteln, in welchen Abschnitten das sich wenigstens näherungsweise wiederholende Muster oder ein davon innerhalb eines Toleranzbereichs abweichendes Muster auftritt.

A first aspect of the present invention relates to a method for analyzing sensor measurement data of a robot or manipulator, comprising:

• - Receiving or determining a data record of sensor measurement data that represent a time course of a measured variable;
• - Identifying at least one at least approximately repeating pattern in the time course of the measured variable in the data set;
• Subdividing the data set into several sections, each section corresponding to a time section which has at least approximately the duration of the at least approximately repeating pattern; and
• Determination of the sections in which the at least approximately repeating pattern or a pattern deviating therefrom within a tolerance range occurs.

With this method, it is possible to determine in which (time) segments the (at least approximately) repeating pattern occurs. This can be useful for controlling or monitoring the robot or manipulator and / or for fault diagnosis. For example, it is possible to specifically examine why the pattern occurs in some time periods and not in others. For example, a correlation between the repeating pattern and certain operating states, movements, positions or the like can be determined. For example, it can be determined that the repetitive pattern always occurs when the manipulator or robot or a certain member of the same is in a certain position or a certain position range, but not when it is not in this position or this position range. This can, if necessary, serve to (early) identify problems which, for whatever reason, are associated with this position or this position range. The same applies, for example, to speed values or speed ranges, etc.

As mentioned above, the data record can be determined while the method is running, for example by evaluating sensor signals that are generated while the robot or manipulator is performing an action or is in a certain state. Alternatively, the data record can also be generated in advance and, if necessary, saved and analyzed at a later point in time. It should also be noted that the actual analysis of the data record does not have to take place in or on the robot or manipulator or in the vicinity of the robot or manipulator, but can be carried out, for example, as part of a remote diagnosis.

“Identifying” a pattern is understood here in particular to mean that the pattern is not known a priori. Rather, in a method according to the invention, a search is made for a pattern or a plurality of patterns, with suitable candidates being determined or selected if necessary. In this context, one also speaks of “learning” a pattern.

When the data set is subdivided, the data of the various sections do not actually have to be separated from one another. Rather will In the present case, subdividing the data set is understood to mean in particular that the data of each section can be examined in the subsequent determination step independently of the data of another section.

In most applications, a certain pattern will not repeat itself exactly, but only within certain limits. Depending on the application, these limits can be set in advance, so that clear criteria can be defined that allow a determination of whether a pattern that occurs is similar to another pattern that occurs with sufficient accuracy so that one can speak of an at least approximately repeating pattern. A tolerance range that is used in the determination step can also be defined by suitable parameters. In the further course of the present description, the term “repeating pattern” is understood to include patterns in which the repetition deviates from the original pattern within a tolerance range, provided the context does not require a different interpretation.

• - Empfangen einer Eingabe, die eine geschätzte Dauer des sich wenigstens näherungsweise wiederholenden Musters angibt und/oder eine Schätzung angibt, wie oft das sich wenigstens näherungsweise wiederholende Muster in dem Datensatz auftritt.

In one embodiment of the method according to the invention, the identification of the at least one at least approximately repeating pattern includes:

• - Receiving an input which indicates an estimated duration of the at least approximately repeating pattern and / or indicates an estimate of how often the at least approximately repeating pattern occurs in the data set.

The input can facilitate a targeted “search” for a repeating pattern.

According to one variant, the input is received from a user via a user interface.

For such an input, the user can, for example, use a graphic representation of the data record (for example in the form of a graphic that shows the time course of the measured variable). Such a graphical representation often reveals the time scales in which the displayed measured variable changes, and based on this the duration of a repeating pattern can often be estimated or it can be estimated how often the repeating pattern occurs in the data set.

Alternatively or additionally, a user can also take into account given technical data of the robot or manipulator. If, for example, it is a robot or manipulator in which rapid movements are to be expected on at least one limb, an axis or a joint, the estimated duration of the repeating pattern can be set relatively short. If only comparatively slower movements are to be expected, the estimated duration can be increased accordingly.

Instead of a (manual) input by a user, the estimated duration can also be determined automatically from the data record. There are several possibilities for this, which are known in principle from the field of data analysis.

According to one embodiment of the method according to the invention, the data record is subdivided using a “sliding window” approach.

A “sliding window” approach is known in principle from the field of data analysis.

According to an embodiment according to the invention, the method further comprises grouping the sections of the data set into at least two segments, so that each segment contains one or more sections, wherein in the event that a segment contains two or more sections, the corresponding time sections are temporally adjacent, wherein all sections in a first of the at least two segments have the at least approximately repeating pattern and all sections in a second of the at least two segments do not have the at least approximately repeating pattern.

Such grouping can provide further information about the occurrence of the repeating pattern. For example, if the pattern occurs in three consecutive periods, then does not appear in the next three time periods, then occurs again in three further periods, then does not appear in the next three periods, etc., a certain periodicity can be identified relatively easily by grouping. Based on this information made clear by the grouping, it is possible, for example, to investigate further why the pattern occurs periodically.

According to a further development, the method also includes the identification of a point in time between two chronologically adjacent segments, in particular the identification of the points in time between all pairs of two chronologically adjacent segments.

In the example of a periodic occurrence of the pattern described above, the points in time after the third, sixth, ninth and twelfth (time) segment, etc. would be identified. Such an identification of these times is also made possible For example, a more targeted search for reasons why the pattern occurs or does not occur after certain points in time. For example, the operating conditions of the robot or manipulator could change at the identified times, which causes the occurrence or absence of the pattern. The identification of the points in time may make it possible or easier to recognize the changes in the operating conditions.

In another example, the occurrence of a pattern in adjacent sections does not alternate periodically with the non-occurrence of the pattern. For example, only a single point in time can be identified in the data record, which can indicate a one-time process (for example a one-time change in an operating condition).

According to one embodiment, the method furthermore comprises adapting a curve, in particular a straight line, to the sensor measurement data within a segment, in particular adapting a straight line in each case to the sensor measurement data within all segments.

By adapting a curve or straight line, the sensor measurement data can be displayed without background noise, for example. Nevertheless, the method according to the invention allows a (better) differentiation between background noise and other factors influencing the sensor measurement data. For example, if a curve or straight line were to be globally adapted to the sensor measurement data, that is to say without taking the various segments into account, it would be assumed that the adaptation is less precise or that the global adaptation does not take into account the different behavior within the segments.

According to one embodiment, the adaptation is carried out using a regression method, in particular using a linear regression method.

Suitable regression methods are known in principle.

According to one embodiment, the method further comprises replacing the identified pattern with a modified pattern in order to determine in which sections the at least approximately repeating pattern or a pattern deviating therefrom within a tolerance range occurs, based on the modified pattern, in particular a further pattern Times to perform.

According to one embodiment, the modified pattern is determined based on the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range.

In particular, the modified pattern can emerge from the first-mentioned pattern by suitable smoothing or filtering, for example in order to filter out high-frequency components that may have been caused by background noise. By using the modified pattern, a better correspondence with the sensor measurement data in the various sections can possibly be determined.

According to one embodiment of the method according to the invention, the modified pattern is determined based on the sensor measurement data that represent the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range, in particular based on this sensor measurement data from several of the sections in which this is at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range occurs, in particular from all sections in which the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range occurs.

According to this embodiment, it can be ensured that a good match between the modified pattern and the sensor measurement data is achieved. In particular if the sensor measurement data from several or all sections in which the pattern occurs are used for this purpose, a sufficiently good or even particularly good match of the modified pattern can be achieved not only with the sensor measurement data from one section, but from several or all sections.

According to one embodiment, the modified pattern can be determined, for example, by averaging based on these sensor measurement data from the several or all sections in which the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range occurs.

By averaging, interfering influences such as background noise can also be filtered out if necessary.

As mentioned above, the method can be performed one more time, based on the modified pattern. The use of the modified pattern may result in a more precise picture, namely firstly with regard to the course of the pattern (for example due to the averaging) and secondly with regard to the appearance of the pattern in the different sections. In other words, after the method has been carried out repeatedly, the (modified) pattern may be recognized in a greater number of sections than was the case when the method was carried out for the first time based on the first-mentioned pattern. The number of sections in which the (modified) pattern occurs may also be reduced. This could be the case, for example, if the pattern was initially determined on the basis of faulty sensor measurement data and too high a number of matches with the (falsified) pattern was detected.

The process can be carried out iteratively, i.e. it can be repeated several times. In in-house tests, the inventors have found that the accuracy can be further improved in this way, i.e. with regard to a more precise determination of the number of correct hits (pattern correctly recognized in a section) and with regard to refining the shape of the pattern.

According to one embodiment of the method according to the invention, the (initial) identification of the at least one at least approximately repeating pattern can be carried out with the aid of an auto-encoder.

A second aspect of the present invention relates to a device for performing one of the methods described above.

A third aspect of the present invention relates to a computer program for performing one of the methods described above.

A fourth aspect of the present invention relates to a computer program product with a program code, which is stored on a medium that can be read by a computer, for carrying out one of the methods described above.

In one embodiment, a computer program product can have, in particular, a non-volatile storage medium for storing a program or with a program stored on it, with the execution of this program causing a system or a controller, in particular a computer, to generate a program described here Process or one or more of its steps.

• 1 zeigt ein erstes Beispiel von Messpunkten, die einen zeitlichen Verlauf einer Messgröße darstellen;
• 2 zeigt ein zweites Beispiel von Messpunkten, die einen zeitlichen Verlauf einer Messgröße darstellen;
• 3 zeigt ein drittes Beispiel von Messpunkten, die einen zeitlichen Verlauf einer Messgröße darstellen;
• 4 zeigt einen Teil der Messpunkte aus 3;
• 5 zeigt die Messpunkte der 5 und zusätzlich eine an diese Messpunkte angepasste Kurve;
• 6 zeigt die an die Messpunkte der 5 angepasste Kurve, aber ohne die Messpunkte; und
• 7 zeigt ein Ablaufdiagramm mit Verfahrensschritten eines beispielhaften erfindungsgemäßen Verfahrens.

Further advantages and features emerge from the subclaims and the exemplary embodiments.

• 1 shows a first example of measuring points which represent a time course of a measured variable;
• 2 shows a second example of measuring points that represent a time course of a measured variable;
• 3 shows a third example of measuring points that represent a time course of a measured variable;
• 4th shows part of the measurement points 3 ;
• 5 shows the measuring points of the 5 and in addition a curve adapted to these measuring points;
• 6th shows the at the measuring points of the 5 adapted curve, but without the measuring points; and
• 7th shows a flow chart with method steps of an exemplary method according to the invention.

1 shows a first example of measuring points that represent a time course (horizontal axis) of a measured variable (vertical axis) in arbitrary units and a measurement curve 1 form. The sensor data on which the measurement points are based can, for example, be position, speed, acceleration, force or torque data of a robot or manipulator, for example measured on a movable member, a joint or an axis of a manipulator or robot. The measurement points shown have an approximately sinusoidal curve, with two low points in the section shown 10 and a high point 11 . The section shown can be viewed as a “sample” within the meaning of the present invention. At the same time, however, only a part of the measuring points can be viewed as a pattern, for example all of the measuring points between the two low points 10 . What is to be regarded as a pattern (also called “shapelet” in the English technical literature) is not specified from the outset and can vary from application to application.

2 shows a second example of measuring points that represent a time course of a measured variable. The pattern 2 initially shows a rise, which then turns into an approximately sinusoidal curve.

3 shows a third example of measuring points that show a time course 3 represent a measured variable. In 3 are also two straight lines 33 and 34 drawn. These are not to be regarded as measuring points and will be explained later.

The measuring points are noisy. In order to eliminate this noise or to obtain a simplified representation, a curve or, in the simplest case, a straight line could in principle be adapted to the measuring points in a manner that is already known. Such a straight line (not drawn in and not to be confused with the straight line 33 and 34 ), for example, the vertical axis (time 0, left edge of the graph) would intersect a little below the horizontal axis and at the end of the period under consideration (time approx. 1500, right edge of the graph) would be slightly above the horizontal axis. The inventors have recognized that, with such a global adaptation to the measurement points, information that may be of interest may be lost.

According to one embodiment according to the present invention, a repetitive pattern is searched for in the temporal course of the measured variable (or in the data set which is illustrated by the measuring points). Such can be recognized in the (time) section 30, for example. This pattern is again separated in 4th and is designated by the reference number 35 Mistake. The pattern 35 has an approximately sinusoidal curve.

In principle, other sections of the time course could also be identified as a pattern or as a candidate for such a pattern. How the identification or selection of patterns or candidates of patterns is carried out will be explained later. But what from 3 is recognizable is the fact that the graphic in the right part (area 32 ) has multiple peaks or "peaks". These repeat themselves at approximately regular intervals. However, the shape of the peaks is not identical, which can be attributed in particular to noise.

After a pattern, such as B. Pattern 35 , has been identified, this can be smoothed if necessary. This results in the (modified) pattern 36 in 6th . 5 shows the pattern 35 and 36 in superimposed representation. The difference between 4th and 5 is difficult to recognize, namely in the area of the lowest point.

The pattern 36 has a length of time 30th , in the 3 is indicated. The history 3 the measurand is now divided into sections of length 30th divided up. This division is not explicitly in 3 shown, but can extend over the entire in 3 area shown extend so that the entire area in adjoining sections of length 30th is divided. It is then determined in which of these sections the pattern 36 occurs. Here, as already indicated, a tolerance range (possibly previously defined) can come into effect, so that non-identical repetitions of the pattern 36 are recognized as repetitions as long as the deviations are within the tolerance range. According to a method according to the invention, it could thus be established, for example, that the pattern 36 in all sections that are in the area 32 lying, occurs while it is found that it is in all sections that are in the area 31 lie, does not occur. This information as such can be of interest in some applications, for example if it would be expected a priori that the pattern would be when operating a robot or manipulator 36 occurs in all periods or does not occur in any period. The information obtained can be used to specifically examine why the pattern is 36 occurs or does not occur in some sections.

According to a further development, the sections of the length 30th are grouped, namely in such a way that all (immediately) consecutive sections in which the pattern 36 occurs, can be combined into one segment. In the in 3 example shown this would be the segment 32 . In addition, (directly) consecutive sections of the length 30th in which the pattern 36 does not occur to another segment - here segment 31 - be summarized. Grouping the sections into different segments can also be used for monitoring or targeted troubleshooting. For example, the length of the segments 31 and 32 can be compared with each other, or it can be examined which operating conditions during the segments 31 or. 32 predominate, which may result in the non-occurrence or occurrence of the pattern 36 cause.

While in 3 only two segments 31 and 32 are shown, it is possible in further examples that several segments are recognizable in a measuring interval in which sections of the length 30th lie in which the pattern 36 occurs or does not occur.

According to a further development, in particular the point in time between two successive segments can be identified. In 3 this would be the point in time approximately at point 400, between the segments 31 and 32 . If more than two segments can be recognized, the times between several or in particular all pairs of successive segments can be identified.

The determination of these points in time can also be used for monitoring or a targeted search for causes that led to the occurrence or non-occurrence of the pattern 36 condition, serve.

According to a further development, within a segment (or within several Segments or all segments) a curve can be adapted to the course of the measured variable. In the simplest case this can be a straight line. In 3 two such straight lines are drawn, namely the straight line 33 and the straight line 34 . Straight 33 is to the sensor measurement data in the segment 31 adjusted, and straight 34 to the sensor measurement data in the segment 32 . You can see that the straight line 34 is essentially parallel to the horizontal axis, with a small positive offset. If, for example, it was to be expected that the straight 34 coincides with the horizontal axis, but this is not the case according to the sensor measurement data, causes for the positive offset can be searched for in a targeted manner. If, on the other hand, a positive offset is to be expected a priori, it can possibly be determined that the sensor measurement data is in the segment 32 meet expectations.

While the straight 34 in the segment 32 runs essentially parallel to the horizontal axis is at the straight line 33 in the segment 31 detect a slope, namely a negative slope. In addition, the straight line lies 33 in the segment 31 in the negative area. Corresponding conclusions can be drawn from this, for example monitoring or searching for causes for the course of the straight line 33 can serve.

Based on the 3 The example explained in particular shows that, according to a method according to the invention, a more differentiated analysis of the sensor measurement data is made possible in comparison to the procedure mentioned above, according to which a curve, in particular a straight line, is globally adapted to all sensor measurement data.

According to a further development of a based on 3 Identifying a repeating pattern can be refined. First, as described above, a repeating pattern, such as. B. the pattern 35 or. 36 , identified and that in 3 shown measuring interval in sections of length 30th divided. Then, as described above, it is determined in which sections the pattern occurs. Then, as an additional step, a modified pattern (or further modified pattern) is determined. When determining the modified pattern, the sensor measurement data are used from at least two, in particular several, in particular all sections in which the originally identified pattern occurs 35 or. 36 was recognized. In particular, the modified pattern can be determined by averaging or the like. The method or some steps of the method are then repeated based on the modified pattern. In particular, it can be determined in which sections of the length 30th the modified pattern occurs before the grouping of the sections into segments as described above is carried out (again). If necessary, by using the modified pattern, a better match of the sensor measurement data in the various sections with the modified pattern can be achieved than when using the originally identified pattern 35 or. 36 was the case. In particular, the use of the modified pattern can have the consequence that the occurrence of the now modified pattern is recognized in more - or also fewer - sections than was the case when the originally identified pattern was used. In this way, false alarms ("false positives" and / or "false negatives") can be avoided or their number reduced.

The steps relating to the refinement of the pattern, that is to say the determination of a modified pattern and the subsequent determination of the sections in which the modified pattern occurs, can optionally be repeated so that an iterative process sequence takes place. The iterative method can optionally be ended after a predetermined number of iterations or in accordance with a termination criterion.

7th shows a flow chart with method steps of an exemplary embodiment of a method according to the invention. This also applies to the 3 to 6th Sensor measurement data or pattern shown is referred to.

After the start 40 the procedure is in one step 41 receive or determine a data record of sensor measurement data. This data record can have been created in advance and, for example, be received in the context of a later analysis by a device that is intended to carry out the method, or, alternatively, it can be determined promptly, ie essentially during the course of the method. The sensor measurement data represent a time course of a measured variable.

Then in one step 42 an at least approximately repeating pattern, for example a pattern 35 or. 36 , identified in the data set in the course of the measured variable over time.

After the pattern has been identified, the data set is divided into several sections, each section corresponding to a time section (for example 30) which is at least approximately the duration of the pattern 35 or. 36 Has.

Finally, in a step 44 it is determined in which sections the pattern or a pattern deviating therefrom within a tolerance range (for example 35 or 36) occurs.

1. 1. Ein Datensatz, der in der Form (y₁,y₂, ···, y_n) vorliegt (oder in diese gebracht werden kann), wobei der Datenvektor y_i ∈ ℝ auch als y_1:n geschrieben werden kann.
2. 2. Ein Maß der Übereinstimmung C(·), das vom Inhalt des Datensatzes abhängt. Dieses Maß der Übereinstimmung kann folgendermaßen ermittelt werden:
   • C(.) = logf (y_i|θ). Dabei ist θ = [β, ε]^T , wobei y_i = i · β + ε und f(·) ist eine Gaußsche radiale Basisfunktion (linearer Gaußscher Kern).
3. 3. Eine Konstante β, die nicht von der Anzahl oder der Position der (Zeit-)Punkte zwischen zwei Abschnitten abhängt. Die Konstante β beschreibt die Granularität der Approximation, also wie genau jeder Abschnitt linear approximiert werden soll. Die Konstante kann beispielsweise vorbestimmt sein oder von einem Benutzer angegeben werden.
4. 4. Eine Konstante K, die die Bedingung C(y_(t+1):s) + C(y_(s+1):T) + K ≤ C(y_(t+1):T) erfüllt. Dabei sind t und T Indizes der Zeit. Außerdem gilt t < s < T. C() ist eine Funktion, die für einen Abschnitt eine Abweichung einer approximierten Linie von einem ursprünglichen Verlauf beschreibt. Für die Konstante K kann beispielsweise 0 gewählt werden.

It will now be described how a pattern or shapelet can be identified in a data set. In this context, an algorithm will first be explained by means of which a data record, which represents the time course of a measured variable, can be divided into suitable sections. The following is used as input for this algorithm:

1. 1. A data set that exists in the form (y ₁ , y ₂ , ···, y _n ) (or can be brought into this), whereby the data vector y _i ∈ ℝ can also be written as y _{1: n} .
2. 2. A measure of correspondence C (·) that depends on the content of the record. This degree of conformity can be determined as follows:
   • C (.) = Logf (y _i | θ). Here θ = [β, ε] ^T , where y _i = i · β + ε and f (·) is a Gaussian radial basis function (linear Gaussian kernel).
3. 3. A constant β that does not depend on the number or the position of the (time) points between two sections. The constant β describes the granularity of the approximation, i.e. how exactly each section should be linearly approximated. The constant can, for example, be predetermined or specified by a user.
4. 4. A constant K that satisfies the condition C (y _{(t + 1): s} ) + C (y _{(s + 1): T} ) + K ≤ C (y _{(t + 1): T} ). Here t and T are indices of time. In addition, t <s <T. C () is a function that describes a deviation of an approximated line from an original course for a section. For example, 0 can be chosen for the constant K.

initialization

1. 1. n die Länge der Daten,
2. 2. F(0) = -β, cp(0) = {}, R₁ = (0),
3. 3. wobei $F\left(s\right)={\text{min}}_{\mathbf{T}\in {\text{T}}_S}\left\{{\displaystyle {\sum }_{i=1}^{m+1}\left[C\left(y_{\left({\tau }_{i-1}+1\right):{\tau }_i}\right)+\beta \right]}\right\}$
   
   Die Funktion K stellt die gesamte Abweichung der approximierten Linie von dem ursprünglichen Verlauf dar.
4. 4. s ist der Zeitindex am Ende des betrachteten Intervalls
5. 5. T_s = {τ: 0 = τ₀ < τ₁ < ... < τ_s = s} ist ein Satz von möglichen (Zeit-)Punkten („change points“) zwischen benachbarten Abschnitten
6. 6. R_i ist bis zum Zeitpunkt i der Satz von ausgewählten bzw. ermittelten Kandidaten für die change points, um später ein „Pruning“ durchzuführen.

Let it be:

1. 1. n is the length of the data,
2. 2. F (0) = -β, cp (0) = {}, R ₁ = (0),
3. 3. where $\mathrm{F.}\left(s\right)={\text{min}}_{\mathbf{T}\in {\text{T}}_{\mathrm{S.}}}\left\{{\displaystyle {\sum }_{i=1}^{m+1}\left[\mathrm{C.}\left(y_{\left({\tau }_{i-1}+1\right):{\tau }_i}\right)+\beta \right]}\right\}$
   
   The function K represents the total deviation of the approximated line from the original course.
4. 4. s is the time index at the end of the considered interval
5. 5. T _s = {τ: 0 = τ ₀ <τ ₁ <... <τ _s = s} is a set of possible (time) points (“change points”) between adjacent sections
6. 6. Up to time i, R _i is the set of selected or determined candidates for the change points in order to carry out a “pruning” later.

iteration

1. 1. Berechnen von F(τ*) = min_{τ∈R τ*} [F(τ) + c(y_{[(τ)] i+1} ]₊₁):τ*) + β],
2. 2. Es sei τ¹ = argmin_{τ∈R τ*} , [F(τ) + C(y(τ _i+1+1):τ*) + β]
3. 3. Es wird gesetzt: cp(τ*) = [cρ(τ¹), τ¹]
4. 4. Es wird gesetzt: R_τ*+1 = {τ ∈ R_τ* U (τ*): F(τ) + C(y_{(τ i+1+1):τ}*) + K ≤ F(τ*)}

The following iteration is carried out for τ * = 1, ···, n:

1. 1. Compute F (τ *) = min _{τ∈R τ *} [F (τ) + c (y _{[(τ)] i + 1} ] ₊₁ ): τ *) + β],
2. 2. Let τ ¹ = argmin _{τ∈R τ *} , [F (τ) + C (y (τ _{i + 1 +1): τ} *) + β]
3. 3. It is set: cp (τ *) = [cρ (τ ¹ ), τ ¹ ]
4. 4. It is set: R _{τ * + 1} = {τ ∈ R _τ * U (τ *): F (τ) + C (y _{(τ i + 1 +1): τ} *) + K ≤ F (τ *)}

The output is a list of change points stored in cp (n).

Using the algorithm described above, a measurement curve can be divided into sections in order to estimate the length of a pattern that occurs or is repeated in the measurement curve. This can then be used in a “sliding window” approach - with such an approach the length of the “window” must be defined.

In simplified terms, the principle on which the above algorithm is based can be described as follows: In a continuous curve, patterns that occur can be viewed as impulses, i.e. as sudden changes in the course. By dividing it into sections, change points are identified. Two successive change points that delimit a section can be viewed as the delimitation of a pattern.

• R. Killick, P. Fearnhead, and I. Eckley, „Optimal detection of changepoints with a linear computational cost,“ Journal of the American Statistical Association, vol. 107, no. 500, pp. 1590-1598, 2012

The algorithm described above is based on an algorithm that comes from the following scientific publication, the content of which is incorporated by reference into this application:

• R. Killick, P. Fearnhead, and I. Eckley, "Optimal detection of changepoints with a linear computational cost," Journal of the American Statistical Association, vol. 107, no. 500, pp. 1590-1598, 2012

It should be noted here that, as mentioned above, the length determined by the algorithm is initially only a (first) estimate of the actual length of the (repetitive) pattern. The estimate can be refined later.

The algorithm described above generally gives a number of sections of different lengths [l _{i, 1} , ..., l _i , _n] , where i is a Is the index designating the i-th section. Statistical analysis can be used to examine which length occurs most frequently. This can be used in the further course of the process as a first estimate of the length of the pattern. In a further development, the k most frequently occurring lengths can also be used as candidates for this and the further procedural steps can be carried out for these candidates. The candidate can then be used as a first estimate of the length of the pattern, through which the best results can be achieved in the further process steps.

• K. Bascol, R. Emonet, E. Fromont and J. Odobez (2016) „Unsupervised Interpretable Pattern Discovery in Time Series Using Autoencoder“

An initial estimate of an at least approximately repeating pattern in different data sets can be carried out with the aid of an autoencoder, the code in the autoencoder representing the repeating patterns. This type of autoencoder is known from the following scientific publication, the content of which is incorporated by reference into this application:

• K. Bascol, R. Emonet, E. Fromont and J. Odobez (2016) "Unsupervised Interpretable Pattern Discovery in Time Series Using Autoencoder"

• J.Grabocka, N. Schilling, M. Wistuba, L. Schmidt-Thieme (2014) „Learning Time-Series Shapelets“, Information Systems and Machine Learning Lab, Universität Hildesheim

The pattern or shapelet can be identified based on a method presented in the following scientific publication, the content of which is incorporated by reference into this application:

• J.Grabocka, N. Schilling, M. Wistuba, L. Schmidt-Thieme (2014) "Learning Time-Series Shapelets", Information Systems and Machine Learning Lab, University of Hildesheim

The input for this method is a data record that shows the time course of a measured variable in a time interval of a predefined length. The output is a probability for different candidates of shapelets that occur in the course of the measured variable over time.

• Zunächst wird das Input-Signal in Abschnitte unterteilt, deren Länge der Länge des Shapelets entspricht. Anschließend wird untersucht, wie sehr der Verlauf der Messkurve innerhalb eines Abschnitts dem Shapelet-Kandidat ähnelt. Hierzu kann innerhalb eines Zeitabschnitts der euklidische Abstand zwischen dem Verlauf der Messkurve von dem Shapelet ermittelt werden. Anschließend wird derjenige Zeitabschnitt ausgewählt, in dem der zeitliche Verlauf der Messgröße dem Shapelet-Kandidat am besten entspricht.
• Sodann wird der kleinste Abstand des Abschnitts mit der besten Übereinstimmung zwischen Messgröße und Shapelet in eine Wahrscheinlichkeit umgewandelt, beispielsweise basierend auf einer Student-t-Verteilung. Dies kann gemäß der folgenden wissenschaftlichen Veröffentlichung erfolgen, deren Inhalt durch Bezugnahme Bestandteil dieser Anmeldung wird:
  • Xie, J., Girshick, R., & Farhadi, A. (2015) „Unsupervised Deep Embedding for Clustering Analysis“ .

For each Shapelet candidate, the following steps are taken:

• First, the input signal is divided into sections, the length of which corresponds to the length of the shapelet. It is then examined how much the course of the measurement curve within a section resembles the Shapelet candidate. For this purpose, the Euclidean distance between the course of the measurement curve from the Shapelet can be determined within a time segment. Then that time segment is selected in which the temporal course of the measured variable corresponds best to the Shapelet candidate.
• The smallest distance of the section with the best agreement between the measured variable and the Shapelet is then converted into a probability, for example based on a Student's t-distribution. This can be done in accordance with the following scientific publication, the content of which is incorporated by reference into this application:
  • Xie, J., Girshick, R., & Farhadi, A. (2015) "Unsupervised Deep Embedding for Clustering Analysis" .

In this context, the Kullback-Leibler divergence can be used as a cost function. This represents a measure of the difference between two probability distributions, in particular between a current distribution and a desired distribution.

In order to select the length of the Shapelet candidates, an input can be requested via a user interface, for example. In this way, a user can, for example, after visually assessing a graphic representation of the course of the measured variable, estimate the (temporal) length of a shapelet appearing in the data set and make a corresponding input. Alternatively, the subdivision into sections can also take place according to the algorithm described above.

In the method according to the invention it can happen that an initially estimated (temporal) length of a pattern is set too long or too short, that is, the actual length of a pattern is shorter or longer than the initially estimated length. In order to determine or estimate the length more precisely or to make a correction, the approach described below can be used.

Let s _k = (s _{k, 1} , ···, s _{k, M} ) be a set of points that represent a repeating pattern of length M. The actual length of the pattern is, however, T, where T <M. During the calculation of a correspondence between a partial sequence of measurement points with the length Q and candidates s _k of patterns, a correlation between the partial sequence and the candidates of patterns is established, taking into account possible shifts (shifts) carried out. A candidate of a pattern taking possible shifts into account can be expressed as follows: $$s_{k,\left(\omega \right)}=\{\begin{array}{cc}\stackrel{\left|\omega \right|}{\overbrace{0,0,\cdots ,0}},s_{k\mathrm{,1}},\cdots ,s{}_{k,\mathrm{M.}-\omega }),& \omega \ge 0\\ s_{k,\mathrm{M.}+\omega },\cdots ,s_{k,\mathrm{M.}},\underset{\left|\omega \right|}{\underbrace{\mathrm{0.0,}\cdots ,0)}}& \omega <0\end{array}$$

Here, ω is the length of a shift, a positive shift being a shift to the right and a negative shift being a shift to the left. This candidate of a pattern, which is subject to a shift, is then correlated with the partial sequence of measurement points t of the partial sequence of length Q. The correlation between the two signals can be defined by the following expression: $$\mathrm{C.}\mathrm{C.}\left(\tau \right)={\displaystyle \sum _{i=\tau }^{\tau +\mathrm{M.}}{t}_i\cdot s_{K,\left(\omega \right),i}}$$

For each candidate s _{k, (ω) of} a pattern that is subjected to a shift, a position can be found in the partial sequence t for which the inner product between the signals becomes a maximum. Then these maxima of the correlation are compared for the various shifts. This results in an optimal shift and thus an optimally determinable or actual length of the repeating pattern.

If the length M of the repeating pattern is less than the actual length T, then at least two patterns can be expected to occur, with the end of one of the patterns near the start of another of the patterns or the end of one with the The beginning of the other pattern coincides, so that one pattern merges into the other pattern. The points of maximum correlation The measurement data that were recorded at different points on a manipulator or robot, for example measured on different axes of a robot, can serve as different “channels”. These can be the same physical quantities (e.g. torque on axis x and torque on axis y) or different (e.g. torque on axis x and (angular) position on axis y). The different channels can also be different physical quantities at the same point (e.g. torque and (angular) position on axis x).

The results obtained for the various channels can then be compared with one another, which - compared with the consideration of only one channel - can be used for more precise control, monitoring or diagnosis.

List of reference symbols

1, 2, 3

Measuring points / curve; Course of a measured variable

10

Low point

11

High point

30th

Time segment / length of a time segment

31, 32

Segment / area

33, 34

Straight

35, 36

Pattern / Shapelet

40-45

Procedural steps

Claims (15)

Method for analyzing sensor measurement data of a robot or manipulator, comprising: - Receiving or determining (41) a data set of sensor measurement data which represent a time course of a measured variable; - Identifying (42) at least one at least approximately repeating pattern (35, 36) over time of the measured variable in the data set; - Subdividing (43) the data set into several sections, each section corresponding to a time section (30) which has at least approximately the duration of the at least approximately repeating pattern; and - Determining (44) in which sections the at least approximately repeating pattern or a pattern deviating therefrom within a tolerance range occurs, the identification of the at least one at least approximately repeating pattern having: Receiving an input which indicates an estimated duration of the at least approximately repeating pattern and / or indicates an estimate of how often the at least approximately repeating pattern occurs in the data set; and wherein the data set is determined during the course of the method by evaluating sensor signals that are generated while the robot or manipulator is performing an action or is in a certain state, or the data set is generated and stored in advance and analyzed at a later point in time . Procedure according to Claim 1 wherein the input is received from a user via a user interface. Method according to one of the preceding claims, wherein the subdivision of the data record takes place using a “sliding window” approach. Method according to one of the preceding claims, further comprising grouping the sections of the data set into at least two segments, so that each segment contains one or more sections, wherein in the event that a segment contains two or more sections, the corresponding time sections are temporally adjacent, wherein all sections in a first of the at least two segments are at least approximately have repeating patterns and all sections in a second of the at least two segments do not have the at least approximately repeating pattern. Procedure according to Claim 4 , further comprising identifying a point in time between two temporally adjacent segments. Procedure according to Claim 4 or 5 , further comprising adapting a curve to the sensor measurement data within a segment. Procedure according to Claim 6 , wherein the adaptation is carried out by means of a regression method. Method according to one of the preceding claims, further comprising replacing the identified pattern with a modified pattern in order to determine in which sections the at least approximately repeating pattern or a pattern deviating therefrom within a tolerance range occurs based on the modified pattern. Procedure according to Claim 8 , wherein the modified pattern is determined based on the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range. Procedure according to Claim 9 wherein the modified pattern is determined based on the sensor measurement data which represent the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range. Procedure according to Claim 10 , wherein the modified pattern is determined by averaging based on this sensor measurement data from the several or all sections in which the at least approximately repeating pattern or the pattern deviating therefrom within a tolerance range occurs. Method according to one of the preceding claims, wherein the identification of the at least one at least approximately repeating pattern is carried out with the aid of an autoencoder. Device for carrying out a method according to one of the preceding claims. Computer program for carrying out a method according to one of the Claims 1 to 12th . Computer program product with a program code, which is stored on a computer-readable medium, for carrying out a method according to one of the Claims 1 to 12th .

Images