Methods for condition monitoring and / or process monitoring of a technical facility

Abstract

The invention relates to a method for condition monitoring and / or process monitoring of a technical device, in particular a machine and / or a plant, wherein a data set to be evaluated is analyzed to detect anomalies in the operation of the technical device. The data set relates to observations of the operation of the technical device, and the analysis is performed to verify whether the analyzed data set relates to an observation of an anomaly in the operation of the technical device. The data set contains values ​​of a plurality of different attributes, each of which relates to a property of the technical device and / or its operation observed within the scope of condition monitoring and / or process monitoring. The attributes include conventional attributes, and the data set is analyzed by a computer using a modified isolation forest algorithm.The algorithm is trained using a set of data records. During training, a forest consisting of a plurality of trees (10) is generated. These trees (10) comprise internal nodes (12, 14, 16), each with exactly two child nodes, and external nodes (18, 20) without child nodes. When generating the trees (10), an internal node (12, 14, 16) is first created as the root node (12) of the respective binary tree (10). During its creation, the root set and / or a portion of the data set, randomly selected for each tree (10) to be generated, is fed to this internal node. When generating the internal nodes (12, 14, 16), the set of data records fed to each internal node (12, 14, 16) is divided into two subsets, and each subset is fed to one of the two child nodes of the respective internal node (12, 14, 16).The internal nodes (12, 14, 16) comprise conventional nodes (14), whose creation involves partitioning into subsets based on a randomly selected conventional attribute. The dataset is evaluated by traversing each binary tree (10) of the forest until it reaches a terminal node. An anomaly value is then calculated for the dataset, inversely related to the average number of conventional nodes (14) that the dataset must traverse in each binary tree (10) before reaching an external node (18, 20). At least one of the internal nodes (12, 14, 16) of each binary tree (10) is a context node (16), whose creation involves partitioning into subsets based on a context attribute.For each of the subsets generated during the creation of a context node (16), a weighting value dependent on the size ratio of the subsets to each other is calculated, and when calculating the anomaly value for the evaluated data set, the weighting value (22) of the subset that was added to the child node of the context node (16) during the creation of the context node (16) is taken into account, to which the evaluated data set is also added when traversing the binary tree (10).

Description

[0001] The invention relates to a method for condition monitoring and / or process monitoring of a technical device according to the preamble of claim 1.

[0002] Condition monitoring, also known as condition monitoring, and / or process monitoring of technical equipment serve to monitor the condition of the equipment and to draw conclusions from the results of the condition monitoring, enabling, for example, predictive maintenance or a proactive response to operational malfunctions. In particular, the failure of parts of the technical equipment, which in many cases is indicated by changes in observed characteristics such as operating parameters, should be detected before it occurs, in order to enable timely shutdown of the equipment and preventive maintenance measures.This allows for the prevention of spontaneous failures during operation, thus avoiding the scenario of a sudden failure during ongoing operations and its potentially serious consequences. Process monitoring makes it possible, in particular, to determine whether quality is satisfactory, production targets are being met, or whether a machine operator has made an error. The monitoring of rapidly wearing parts, such as tools, can also be assigned to this area as an alternative to condition monitoring.

[0003] The technical equipment in question can be, in particular, a machine and / or a system. Such technical equipment is often complex in practice. Therefore, it is difficult in practice to identify and predict a specific scenario early on based on measured values.

[0004] To manage the multitude of possible scenarios, condition monitoring and / or process monitoring procedures often involve data analysis for anomaly detection. This utilizes machine learning or artificial intelligence methods. These are typically trained on normal behavior so they can recognize deviations from it.

[0005] One such algorithm is the Isolation Forest. The Isolation Forest algorithm is based on the assumption that data sets relating to rare events, such as anomalies, can be more easily isolated than other data sets relating to the far more frequent normal operation of the technical equipment. This algorithm can be described using the structure of a binary tree. The procedure involves creating a forest of trees during the algorithm's training. The data set to be evaluated is processed by traversing the trees of this forest. An anomaly value can then be calculated for the data set.

[0006] In practice, however, certain properties of technical equipment observed within such procedures can assume varying values ​​without this resulting from an event relevant for condition monitoring and / or process monitoring in the form of an anomaly.

[0007] For example, the speed at which a production process operates can vary depending on capacity utilization, i.e., depending on external influences. This can directly affect the rotational speeds at which the drives of the technical equipment used in the production process are operated. If production typically takes place at speeds within a certain, relatively narrowly defined range, the resulting fluctuations in the values ​​of observed properties of the technical equipment will be comparatively small. However, if on a particular day—for example, due to an external event—there is a significant underutilization and the production speed is therefore drastically reduced for a short period, individual data sets will be generated during this time window in which the values ​​of a number of observed properties deviate significantly from the values ​​that occur during normal operation.A classic Isolation Forest algorithm would identify these datasets as conspicuous anomalies.

[0008] Furthermore, the application of the algorithm may introduce distortions if, for example, production speed is generally lower on weekends than during the week, perhaps due to reduced shifts. Observations resulting from weekend operation of the technical equipment will tend to show higher anomaly values ​​than observations made during the week.

[0009] This could be addressed by specifically training the algorithm for the different distinguishable cases. This might still be practically feasible when there is a clear distinction between production on weekends and on weekdays. However, in cases where a multitude of different operating scenarios each influence the values ​​of the observed properties, such targeted training of the algorithm for different scenarios is no longer practical, especially since this would require an enormous amount of training data.

[0010] In practice, this significantly limits the efficiency and benefit of using the Isolation Forest algorithm to detect anomalies in connection with a condition monitoring and / or process monitoring procedure.

[0011] The invention is therefore based on the objective of demonstrating a method for condition monitoring and / or process monitoring of a technical device in which the aforementioned disadvantages do not occur or at least only to a reduced extent.

[0012] The problem is solved by a method for condition monitoring and / or process monitoring of a technical device with the features of independent claim 1. The features of the dependent claims relate to advantageous embodiments.

[0013] The method is used for condition monitoring and / or process monitoring of a technical system. The method specifically involves conducting observations of the system's operation, particularly automated observations, and generating data sets for each observation, also automated, which each record pertains to a specific aspect of the system's operation.

[0014] The technical equipment is, in particular, a machine and / or a system. The technical equipment may, in particular, include an electric drive system.

[0015] To detect anomalies in the operation of the technical equipment, a data set is analyzed. This data set relates to observations of the operation of the technical equipment.

[0016] The observation of the operation of the technical equipment involves, in particular, the observation of a plurality of properties of the technical equipment and / or a plurality of properties of the operation of the technical equipment, in particular whereby the observation takes place at a specific time and / or in a specific period of time.

[0017] Observing a property primarily involves measuring its value, for example, using a sensor. Alternatively and / or additionally, observation can also involve reading a property's value, particularly from a component of an automation system, such as a control unit.

[0018] The evaluation is carried out to verify whether the analyzed data set relates to an observed anomaly in the operation of the technical equipment. This enables further measures to be taken upon detection of an anomaly, in particular more in-depth troubleshooting and / or intervention in the operation of the affected technical equipment, such as its shutdown.

[0019] The dataset contains values ​​for multiple different attributes. Each attribute relates to a property of the technical equipment observed during condition monitoring and / or process monitoring. The attributes can be numerical and / or categorizing. Numerical attributes are those that can assume a numerical value, such as an attribute relating to the magnitude of an electrical voltage measured by a sensor. Categorizing attributes are those that can only assume a fixed set of different states. A categorizing attribute could, for example, be the day of the week on which an observation was made.

[0020] The attributes include conventional attributes. Conventional attributes are, in particular, attributes that are suspected and / or known to be subject to changes in the event of an anomaly.

[0021] In this process, the algorithm is trained using a basic set of data records, similar to a conventional Isolation Forest algorithm. During training, a forest consisting of multiple trees is generated, again similar to a conventional Isolation Forest algorithm. These trees can be binary trees.

[0022] The dataset is evaluated by a computer using a modified Isolation Forest algorithm. The dataset is analyzed by traversing the trees of the forest.

[0023] The basic set of data records contains, in particular, data records relating to observations of a specific technical device at different times and / or over different periods. Alternatively and / or additionally, the procedure may provide that the basic set of data records contains observations relating to the operation of other technical devices. In this way, the available basic set of data records can be increased.

[0024] The other technical equipment may be similar and / or identical. Similar technical equipment is understood to mean, in particular, equipment that differs from the technical equipment being monitored for condition monitoring and / or process monitoring only with respect to those technical characteristics that have no or a negligible influence on the behavior of the observed properties. A negligible influence is understood to mean an influence that, while measurable, would not significantly affect the anomaly detection results even when using a conventional isolation forest algorithm.

[0025] For example, a different paint color on a housing will generally have no effect on observed properties. However, a slightly different housing design can measurably affect individual observed properties, for instance, if it slightly alters the vibration behavior of the technical equipment and this has a minor impact on observed vibrations. Such an influence can be measurable, but as long as it does not significantly affect the anomaly detection results, it can still be considered a similar technical device for the purposes of this procedure.

[0026] Alternatively and / or additionally, the other technical facilities can also be of a different type. A significant advantage of the described method is that, due to the use of context nodes and context attributes, even technically different facilities can be used to obtain data sets for expanding the database, such that no meaningful results can be achieved using a conventional isolation forest algorithm.

[0027] The trees consist of internal nodes, each with exactly two child nodes, and external nodes without child nodes. The external nodes of such trees are also called leaves.

[0028] When generating the trees, an internal node is first created as the root node of the respective binary tree. Upon its creation, the root node is fed the entire set and / or a portion of the data set. Specifically, the portion of the data set is randomly selected for each tree to be generated.

[0029] When creating internal nodes, the set of data records fed to each internal node is divided into two subsets, and each of the two subsets is fed to one of the two child nodes of that internal node. These subsets are, in particular, discrete subsets.

[0030] The procedure specifically stipulates that an external node is created when a subset containing at most one data record is passed to the node during its creation. If a subset contains only one data record or is an empty set, further subdivision of this subset is not possible. Accordingly, it is advantageous for the node to which this subset is passed during the creation of the binary tree to be an external node.

[0031] The procedure can provide that internal nodes are always created when generating the trees if the set of data records supplied to the respective node during its creation contains at least two data records.

[0032] The internal nodes comprise conventional nodes, the creation of which involves the division into subsets based on a randomly selected conventional attribute, as is the case, in particular, when applying a conventional Isolation Forest algorithm.

[0033] For the dataset being evaluated, an anomaly value is calculated that is inversely related to the average number of conventional nodes that the dataset must traverse in each binary tree when traversing the trees of the forest before reaching an external node. In other words, the anomaly value is inversely related to the average length of the path from the root node to the external node that the dataset reaches when traversing the respective tree.

[0034] The data set traverses the trees, in particular as in a conventional Isolation Forest algorithm, in particular such that after passing through an internal node, the data set is fed to the child node of that internal node to which, when creating that internal node, the subset in which the data set would have been assigned when creating that internal node was also fed.

[0035] The problem is solved in particular by ensuring that at least one of the internal nodes of each binary tree is a context node, the creation of which involves the division into subsets based on a context attribute. Context attributes are, in particular, attributes that are suspected and / or known to be subject to changes independently of the occurrence of an anomaly.

[0036] In particular, the context attributes are numerical attributes whose values ​​are suspected and / or known to fluctuate even during the normal operation of the technical equipment. This fluctuation can result from external influences on the equipment, for example, if the observed property related to the context attribute is an ambient temperature. Alternatively and / or additionally, the fluctuation can result from intentional manipulation of the equipment, for example, if the speed of a production process is increased and / or decreased. In such cases, the result is a significant change in the value of the affected context attribute, without any anomaly being present.

[0037] Alternatively and / or additionally, the context attributes are primarily categorical attributes where it is suspected and / or known that a change in category also entails a change in the normal behavior of the technical equipment. For example, the category "weekday" might apply if a technical device operates with reduced load on weekends and holidays.

[0038] The procedure further stipulates that for each of the subsets generated during the creation of a context node, a weight value dependent on the relative sizes of the subsets is calculated. When calculating the anomaly value for the evaluated data set, the weight value of the subset that was assigned to the child node of the context node during its creation—the same child node to which the evaluated data set is also assigned when traversing the binary tree—is taken into account. Similarly, the procedure can provide that when calculating the anomaly value of a data set that traverses multiple context nodes during a binary tree, a corresponding multiple weight values ​​are also considered.

[0039] This approach is based on the understanding that the observed properties of a technical device are typically interconnected in complex ways. For example, if identical technical devices are operated at different ambient temperatures, this can affect a whole range of observed properties.For example, a scenario may occur where a majority of the technical equipment is operated at similar and generally higher temperatures, while a few technical equipment units, or even a single technical equipment unit, is operated at a significantly lower temperature. Assuming that the basic set of data sets used for the procedure contains the same number of data sets for each technical equipment unit, then when applying a conventional Isolation Forest algorithm, it can be expected that data sets relating to observations of the technical equipment unit operated at a lower temperature will tend to calculate higher anomaly values ​​than data sets relating to observations of the technical equipment unit operated at a higher temperature.

[0040] The weighting value makes it possible to compensate for such effects. Using the weighting value, the fact that, due to the complex relationships between the properties of the real technical setup, the context attribute—according to which a set of data records in a context node is divided into two subsets—is highly likely to exhibit correlations with conventional attributes, leading to the "distortion" of the anomaly values ​​described above, can be compensated for.

[0041] The procedure can involve dividing a set into two subsets based on an attribute by randomly setting a threshold. Records whose values ​​for this attribute exceed this threshold are then moved to one subset, and records whose values ​​fall below this threshold are moved to the other. This can apply to conventional nodes and / or context nodes. If the threshold is chosen randomly, records with "normal" values ​​are more likely to be isolated or assigned to the smaller of the two subsets than records with "normal" values. Applied to context nodes, the random selection of the threshold, due to the fact that an entire forest is being considered, leads to records with rare values ​​of context attributes being more likely to be assigned to the smaller of the two subsets than records with more frequent values ​​of context attributes.Averaged across the entire forest, weighting values ​​are then assigned to the data sets with the rare values ​​of the context attributes, the consideration of which in the calculation of the anomaly value leads to its reduction.

[0042] The procedure can stipulate that the random threshold is only set within a specific interval of possible values ​​for the respective attribute. For example, the procedure can specify that the threshold is set within an interval that lies between certain quantiles of the distribution of the values ​​of the respective attribute in the set of records to be partitioned. These could be, for example, the 25th percentile and the 75th percentile. The use of such intervals in general, and the aforementioned quantiles in particular, has proven useful in practice to prevent an unfavorable choice of threshold from resulting in all records in an internal node being assigned to one of the two subsets.

[0043] The procedure can provide that, when an internal node is created and the distribution of the data records fed to that context node is based on a categorizing attribute, the categories are mapped to numerical values. Such a mapping of categories to numerical values ​​allows the categorized attributes to be treated analogously to numerical attributes with regard to distribution based on exceeding or falling below a threshold.

[0044] The method can specify that the root nodes of the trees are context nodes. It has been shown that the method yields particularly good results regarding anomaly detection when the root node is a context node.

[0045] The procedure can provide that the trees are generated in such a way that every node except the root node of the respective binary tree has exactly one parent node.

[0046] The method can be configured to generate trees in such a way that, when creating a binary tree at a predetermined depth, only external nodes are created. In other words, the method can specify the depth of the lowest level of the binary tree, and the nodes at that level are created only as external nodes. In this way, the depth of the trees can be limited in practice.

[0047] The method can, particularly when a specified depth of the trees to be generated, provide that internal nodes are always created when generating the trees if the set of data records fed to the respective node during its creation contains at least two data records and the specified depth of the binary tree has not yet been reached at the time of the respective node's creation. In other words, any amount of data fed to a node containing at least two data records can lead to the creation of an internal node, unless the node to be created is located at the lowest level of the binary tree.

[0048] The procedure can provide that when an external node is created, in particular when an external node of the lowest level of the respective binary tree is created, a node value is assigned to this external node, which depends on the number of data records in the subset that is supplied to this node when it is created.

[0049] The method can provide that the node value of the external node, which the data set reaches when traversing a binary tree during evaluation, is included in the calculation of the data set's anomaly value. In this way, a loss of accuracy resulting from the limitation of the possible tree depth can be at least partially compensated. In this context, the method can specifically provide that external nodes at the lowest level of a binary tree, which are fed a larger subset of data sets during their creation, receive a node value whose inclusion in the calculation of the anomaly value leads to a lower anomaly value, while external nodes at the lowest level of a binary tree, which are fed a smaller subset of data sets during their creation, receive a node value whose inclusion in the calculation of the anomaly value leads to a higher anomaly value.

[0050] The method can, in particular, provide that the trees are generated in such a way that, up to a predetermined depth, the internal nodes of the trees are context nodes. This means that the internal nodes through which the dataset passes during evaluation in a given binary tree are initially only context nodes. It has been shown that such a tree structure leads to particularly advantageous results in anomaly detection in practice.

[0051] The method can, in particular, provide that the trees are generated in such a way that, from a predetermined depth onward, the internal nodes of the trees are conventional nodes. This means that, from the point at which the dataset is processed in a given binary tree, the internal nodes are all conventional nodes. It has been shown that such a tree structure leads to particularly advantageous results in anomaly detection in practice.

[0052] The method can, in particular, specify that the trees are generated such that all internal nodes located on the same level of the binary tree are either context nodes or conventional nodes. The number of levels containing context nodes can, in particular, be at least 30% and / or at most 50% of the number of levels in the respective binary tree. Empirical evidence has shown that such a tree structure leads to particularly advantageous results in anomaly detection.

[0053] The procedure may stipulate that the observed properties are operating variables, operating parameters, structural characteristics, environmental variables and / or environmental parameters of the technical equipment and / or a product parameter.

[0054] The procedure may stipulate that the conventional attributes include at least one conventional attribute selected from: - Raw data from a drive train of the technical equipment, in particular raw data of current, voltage, power, vibration, acoustics, rotational speed, position, torque and / or quantities derived from these quantities, - Raw data from the operation of the technical equipment, in particular variables from control devices such as PLC controllers and / or frequency converters, and / or - data calculated from these raw data, in particular during the processing of these for the purpose of carrying out the procedure, in particular where the calculated data are statistical parameters,

[0055] Includes signal transformations and / or neural networks for processing.

[0056] The procedure may stipulate that the context attributes include at least one context attribute selected from: - Raw data from environmental sensors of the technical equipment, in particular raw data of temperature, humidity, date and / or time, ambient vibration and / or ambient acoustics, - Raw data from processes, controls and / or drives, in particular a state from a state machine of a control and / or a drive, raw data from current process steps, a process setpoint, in particular from a recipe, operator inputs, system inputs via interfaces, a state of a drive, a motion setpoint from a drive, - calculated data from these raw data, in particular during the processing of these for the purpose of carrying out the procedure and / or - Configuration data relating to the configuration of the technical equipment, in particular a software version, a hardware version, a mechanical dimension, a variant configuration, a software parameter.

[0057] The method may, in particular, provide that the computer recognizes anomalies based on the anomaly value of the evaluated data set, determines whether the observation underlying this data set is likely to be an anomaly, and generates a notification to a user. The generation of the notification may be automated, particularly by the computer. The method may also provide that the notification contains information enabling the identification of the data set, the identification of the technical equipment, and / or the identification of the time and / or time period in which the observation was carried out.

[0058] Further practical embodiments and advantages of the invention are described below in connection with the drawing. Fig. Figure 1 shows a schematic representation of an exemplary binary tree of an exemplary procedure.

[0059] Fig. Figure 1 shows a schematic representation of an exemplary binary tree 10 of an exemplary procedure for condition monitoring and / or process monitoring of a technical facility. In carrying out the exemplary procedure, a data set relating to an observation of the operation of the technical facility is evaluated to detect anomalies in the operation of the technical facility. The evaluation is performed to check whether the evaluated data set relates to an observation of an anomaly.

[0060] The dataset to be evaluated is analyzed using a modified Isolation Forest algorithm, in which the dataset iterates through the trees (10) of a forest. During the initial training of the algorithm, a forest consisting of a plurality of trees (10), such as one in Fig. 1 is shown as an example, generated.

[0061] The trees 10 comprise internal nodes 12, 14, and 16, each with exactly two child nodes, and external nodes 18 and 20, which have no child nodes. When creating the trees 10, an internal node 12 is first created as the root node 12 of the respective binary tree 10. A set of data records is fed to the root node 12 upon its creation.

[0062] In the exemplary procedure for generating internal nodes 12, 14, 16, the set of data records supplied to each internal node is divided into two subsets, and each of the two subsets is supplied to one of the two child nodes of the respective internal node 12, 14, 16. In the example shown, the subsets are discrete subsets. Fig. In each node (12, 14, 16, 18, 20), the number 1 represents an example of the number of data records assigned to that node upon its creation. For internal nodes 12, 14, 16, the number of data records assigned to each node is shown in the bottom row.

[0063] The procedure, as in the example shown, can provide that an external node 18 is created when node 18 is assigned a subset containing at most one data record during its creation. The internal nodes 12, 14, and 16 comprise conventional nodes 14, whose creation is based on a randomly selected conventional attribute. At least one of the internal nodes 12, 14, and 16 of each binary tree 10 is, as in the example shown, a context node 12 or 16, whose creation is based on a context attribute. As in the example shown, the root node 12 can also be a context node 12.

[0064] In the case of the example binary tree 10, the first row of each representation of internal nodes 12, 14, 16 specifies the attribute according to which the set of data records fed to these nodes 12, 14, 16 is distributed when these nodes are created. In the example shown, attributes 1, 2, and 3 are context attributes, while attributes 4, 5, 6, and 7 are conventional attributes.

[0065] The procedure further stipulates that for each of the subsets generated during the creation of a context node 12, 16, as in Fig. Figure 1 shows a weighting value 22 that is calculated depending on the size ratio of the subsets to each other.

[0066] As in the exemplary procedure shown, the trees 10 can be generated such that each of the nodes 14, 16, 18, 20, with the exception of the root node 12, has exactly one parent node. The parent nodes are, in particular, internal nodes 12, 14, 16.

[0067] The method, as in the example shown, can be designed so that the trees 10 are generated in such a way that, when creating a binary tree 10 at a predefined depth of the binary tree 10, only external nodes 20 are created. The depth can be specified in terms of levels of the binary tree. In the example method, the trees 10 have a depth of 5 levels, as shown.

[0068] As in the example shown, the procedure can stipulate that when generating the trees 10, internal nodes 12, 14, 16 are always created if the set of data records fed to the respective node 12, 14, 16 during its creation contains at least 2 data records and the specified depth of the binary tree has not yet been reached at the time of the respective node's creation. The latter applies in the illustrated example for the upper 4 levels of the exemplary binary tree 10.

[0069] As illustrated by example, the procedure can stipulate that when an external node 18, 20 is created, this external node is assigned a node value that depends on the number of data records in the subset that are fed to this node 18, 20 during its creation. In the exemplary procedure shown, the node value corresponds to the number of data records fed to this node 18, 20 during its creation.

[0070] As in the example shown, the procedure can provide that the trees 10 are generated in such a way that, up to a predetermined depth of the trees 10, the internal nodes 12, 16 of the trees 10 are context nodes 12, 16. This is the case in the example shown for the top two levels of the exemplary binary tree 10.

[0071] As illustrated by example, the procedure can provide that the trees 10 are generated in such a way that, from a given depth of the trees 10 onwards, the internal nodes 14 of the trees are conventional nodes 14. This is the case in the example shown for the lowest three levels of the exemplary binary tree 10.

[0072] For each dataset to be evaluated, an anomaly value is calculated that is inversely related to the average number of conventional nodes that the dataset must traverse in each binary tree when traversing the trees of the forest before reaching an external node. In other words, the anomaly value is inversely related to the average length of the path from the root node to the external node that the dataset reaches when traversing the respective tree.

[0073] The data set traverses the trees, in particular as in a conventional Isolation Forest algorithm, in particular such that after passing through an internal node, the data set is fed to the child node of that internal node to which, when creating that internal node, the subset into which the data set would have been divided when creating that internal node was also fed.

[0074] During the evaluation of the data set to be analyzed, the data set traverses the trees 10 of the forest until it reaches an external node 18, 20 of the respective binary tree 10. An anomaly value is calculated for the data set to be analyzed, which is inversely related to the number of conventional nodes 14 that this data set must traverse on average in each binary tree 10 when traversing the trees 10 of the forest until it reaches an external node 18, 20.

[0075] The data set traverses the trees 10 in such a way that, after passing through an internal node 12, 14, 16, the data set is fed to that child node of this internal node 12, 14, 16 to which, when creating this internal node 12, 14, 16, the subset into which the data set would have been divided when creating this internal node 12, 14, 16 was also fed.

[0076] The exemplary procedure stipulates that, when calculating the anomaly value for the evaluated data set, the weight value 22 of the subset is taken into account that subset which, during the creation of context node 12, 16, was assigned to the child node of context node 12, 16 to which the evaluated data set is also assigned when traversing the binary tree 10. As in the example shown, the procedure can provide that, when calculating the anomaly value of a data set that traverses multiple context nodes 12, 16 while traversing a binary tree 10, a corresponding multiple weight values ​​are also taken into account. In the case of the example binary tree 10, a data set to be evaluated will always traverse two context nodes 12, 16, so that, accordingly, two weight values ​​22 are considered when calculating the anomaly value.

[0077] The procedure can further provide that, as in the example shown, the node value of the external node 18, 20, which the data set reaches when traversing a binary tree 10 during the evaluation of the data set, is included in the calculation of the anomaly value of the data set.

[0078] The features of the invention disclosed in this description, in the drawings, and in the claims can be essential for realizing the invention in its various embodiments, both individually and in any combination. The invention is not limited to the described embodiments. It can be varied within the scope of the claims and taking into account the knowledge of the person skilled in the art. Reference symbol list 10 Binary tree 12 root nodes (internal nodes / context nodes) 14 conventional nodes (internal nodes) 16 context nodes (internal nodes) 18 external nodes 20 external nodes 22 Weighting value

Claims

Method for condition monitoring and / or process monitoring of a technical device, in particular a machine and / or a plant, wherein a data set to be evaluated is analyzed to detect anomalies in the operation of the technical device, wherein the data set relates to an observation of the operation of the technical device, wherein the analysis is carried out to check whether the analyzed data set relates to an observation of an anomaly in the operation of the technical device, wherein the data set contains values ​​of a plurality of different attributes, wherein the attributes each relate to a property of the technical device and / or the operation of the technical device observed within the framework of condition monitoring and / or process monitoring, wherein the attributes include conventional attributes, wherein the data set is analyzed by a computer using a modified isolation forest algorithm.wherein the algorithm is trained using a basic set of data records, wherein during the training of the algorithm a forest of a plurality of trees (10) is generated, wherein the trees (10) comprise internal nodes (12, 14, 16) each with exactly two child nodes and external nodes (18, 20) without child nodes, wherein during the generation of the trees (10) an internal node (12, 14, 16) is first generated as the root node (12) of the respective binary tree (10), to which the basic set and / or a portion of the basic set of data records, in particular randomly selected for each tree (10) to be generated, wherein during the generation of the internal nodes (12, 14, 16) the set of data records supplied to the respective internal node (12, 14, 16) is divided into two subsets and each of the two subsets is supplied to one of the two child nodes of the respective internal node (12, 14, 16). will be, where the internal nodes (12, 14,16) comprise conventional nodes (14) whose generation involves the division into subsets based on a randomly selected conventional attribute, wherein the data set is evaluated by traversing each binary tree (10) of the forest until it reaches a terminal node and calculating an anomaly value for the data set that is inversely related to the average number of conventional nodes (14) that this data set must traverse in each binary tree (10) when traversing the trees (10) of the forest until it reaches an external node (18, 20), characterized in that at least one of the internal nodes (12, 14, 16) of a respective binary tree (10) is a context node (16) whose generation involves the division into subsets based on a context attribute.wherein for each of the subsets generated during the creation of a context node (16) a weighting value is calculated which depends on the size ratio of the subsets to each other and wherein, when calculating the anomaly value for the evaluated data set, the weighting value (22) of the subset which was added to the child node of the context node (16) during the creation of the context node (16) to which the evaluated data set is also added when traversing the binary tree (10) is taken into account. Method according to claim 1, characterized in that when dividing a set into two subsets based on an attribute, a random limit value is set and the data records whose values ​​of this attribute exceed this limit value are transferred to one subset and the data records which fall below this limit value are transferred to the other subset. Method according to claim 1 or 2, characterized in that the root nodes (12) of the trees are context nodes (16). Method according to one of the preceding claims, characterized in that the trees (10) are generated such that up to a predetermined depth of the tree (10) all generated internal nodes (12, 14, 16) are generated as context nodes (16). Method according to one of the preceding claims, characterized in that the trees (10) are generated such that each node of the tree (10) is either an internal node (12, 14, 16) with exactly two child nodes or an external node (18, 20) without child nodes. Method according to one of the preceding claims, characterized in that the trees (10) are generated such that each node except the root node (12) of the respective binary tree has exactly one parent node. Method according to one of the preceding claims, characterized in that the trees (10) are generated such that when generating a binary tree (10) at a predetermined depth of the binary tree (10) only external nodes (18, 20) are generated. Method according to one of the preceding claims, characterized in that when generating an external node (18, 20), in particular when generating an external node (18, 20) of the lowest level of the respective binary tree (10), a node value is assigned to this external node (18, 20) which depends on the number of data records in the subset that is supplied to this node during its generation, in particular wherein the node value of the external node (18, 20) that the data record reaches when traversing a binary tree (10) during the evaluation of the data record is included in the calculation of the anomaly value of the data record. Method according to one of the preceding claims, characterized in that the trees (10) are generated such that up to a predetermined depth of the trees (10) the internal nodes (12, 14, 16) of the trees (10) are context nodes (16) and / or that from a predetermined depth of the trees (10) the internal nodes (12, 14, 16) of the trees (10) are conventional nodes (14). Method according to one of the preceding claims, characterized in that the trees (10) are generated such that all internal nodes (12, 14, 16) that are on the same level of the binary tree (10) are either context nodes (16) or conventional nodes (14), in particular wherein the number of levels containing context nodes (16) is at least 30% and / or at most 50% of the number of levels of the respective binary tree (10). Method according to one of the preceding claims, characterized in that the observed properties are operating variables, operating parameters, structural properties, environmental variables and / or environmental parameters of the technical equipment and / or a product parameter. A method according to one of the preceding claims, characterized in that the conventional attributes comprise at least one conventional attribute selected from: - raw data from a drive train of the technical equipment, in particular raw data of a current, voltage, power, vibration, acoustics, rotational speed, position, torque and / or quantities derived from these quantities, - raw data from operation of the technical equipment, in particular variables from control devices such as PLC controllers and / or frequency converters and / or - data calculated from these raw data, in particular within the scope of processing them for the purpose of carrying out the method, in particular wherein the calculated data include statistical parameters, signal transformations and / or neural networks for processing. A method according to one of the preceding claims, characterized in that the context attributes comprise at least one context attribute selected from: - raw data from environmental sensors of the technical equipment, in particular raw data of temperature, humidity, date and / or time, ambient vibration and / or ambient acoustics, - raw data from processes, controls and / or drives, in particular a state from a state machine of a control and / or a drive, raw data from current process steps, a process setpoint, in particular from a recipe, operator inputs, system inputs via interfaces, a state of a drive, a motion setpoint from a drive, - data calculated from these raw data, in particular within the scope of processing them for the purpose of carrying out the method, and / or - configuration data relating to the configuration of the technical equipment.in particular a software version, a hardware version, a mechanical dimension, a variant configuration, a software parameter. Method according to one of the preceding claims, characterized in that the observation of a property involves measuring a value of a property and / or reading out a value of a property. A method according to one of the preceding claims, characterized in that the computer recognizes anomalies based on the anomaly value of the evaluated data set, determines whether the observation underlying this data set is likely to be an anomaly, and automatically generates a notification to a user, in particular wherein the notification includes information that enables the identification of the data set, the identification of the technical equipment and / or the identification of the time and / or time period of the observation.

Images