Monitoring device for monitoring state of machine

Abstract

The invention relates to a monitoring device for monitoring the state of a machine (10), comprising (20):-a machine learning unit (21) designed to: receive sensor data, which are collected on the machine (10) during operation, comprising at least one parameter of the machine (10); and determining an abnormal result and at least one feature in the sensor data that demonstrates the abnormal result from the sensor data, and transferring the abnormal result and the feature to a transmission unit (22); -a transmission unit (22) which is designed to determine at least one indication within the sensor data for the outcome of the anomaly as a function of the at least one feature and to deliver the at least one indication to an interpretation unit (23); and-the interpretation unit (23), which is designed to check the exception result as a function of the at least one indication and a rule-containing criterion, and to output a control command for execution on the machine if the exception result is affirmative, an exception is confirmed, the criterion comprises at least one assignment from an indication to a fault type of the machine (10).

Description

Technical Field

[0001] The invention relates to a monitoring device for monitoring the state of a machine, as well as a corresponding method and a computer program product. Background Art

[0002] The monitoring of industrial systems, whether for the purpose of identifying anomalies or for the purpose of predicting maintenance cycles (predictive maintenance), is based on an optimal understanding of the physical and technical context of the system and the ability to describe it analytically, or on the availability of a sufficient amount of good and bad data to train artificial intelligence (AI) in the form of machine learning models, especially classifiers, well enough.

[0003] Typical requirements for monitoring systems here are cold start capability, i.e. semi-automatic fault detection of the system to be monitored, even when there is no data or a small amount of data, or a conclusive fault identification is required, which goes far beyond a warning in the form of "Attention!" The decision about the existence of a fault should be traceable, and the solution should be able to be quickly and flexibly transferred to "similar but not identical" devices or equipment.

[0004] The main problems here are: technical understanding is insufficient for an accurate and comprehensive analytical description, and there is not enough data to train the ML models before going into operation. Expert knowledge often exists as "gut feelings" and experience or in the form of generic descriptions, which are difficult to compare with real sensor values. In addition, detailed knowledge of the failure processes and failure behaviors of system types is not sufficient.

[0005] Simulation models can provide a certain description of the system to be monitored and its behavior, but rarely model all interrelated effects, such as multi-physics methods, statistically weakly correlated effects, and time offsets between faults and failures. In addition, obtaining quantitative conclusions with the help of simulation models is very complex, because the model must be accurately parameterized for each industrial system.

[0006] The system or production process to be monitored is usually equipped with multiple sensors of different types, which provide data of different qualities, such as signal quality, and are fed with different types of data streams. For example, the signal is transmitted as a voltage between 0...1V or 0...10V or -5...5V or a current of 0...24mA, or semantically annotated as ZAHL+EINHEIT+FEHLERMASS+METAINFO

[0007] (NUMBER+UNIT+DEGREE OF ERROR+METAINFO). Sensor data can also be implemented in different data modalities, such as time series, RGB-2D images, 3D terrain data, multi-dimensional spectral camera data, sporadic single measurements without a fixed period, event-based measurements, etc. Summary of the invention

[0008] The object of the present invention is therefore to provide a device which performs semi-automatic and retroactive fault detection when little data is available from the system and outputs meaningful fault indications and which can also be used for similar but not identical systems.

[0009] This object is achieved by the measures described in the independent claim. Advantageous developments of the invention are presented in the dependent claims.

[0010] According to a first aspect, the invention relates to a device for monitoring the status of a machine, the device comprising:

[0011] a machine learning unit that receives sensor data collected on the machine during operation, including at least one parameter of the machine, and determines an abnormal result and at least one feature in the sensor data that demonstrates the abnormal result based on the sensor data, and transmits the abnormal result and the feature to a transmission unit;

[0012] - the transmission unit determines at least one indication within the sensor data for the abnormal result based on the at least one feature and forwards the at least one indication to the interpretation unit; and

[0013] The interpretation unit checks the exception result based on the at least one indication and a criterion containing rules, and outputs a control instruction for execution on the machine if the exception result is positive, confirming the exception, the criterion including at least one assignment from the indication to a fault type of the machine.

[0014] The device is modularly divided into a data-based machine learning unit and a "physics-based" interpretation unit. The former is based on machine learning, i.e., is designed using a data-driven approach. Domain knowledge about the machines to be monitored is not required or is only minimally required here. Domain knowledge, including possible anomalies, is stored in an abstract manner via criteria. Storage in the form of abstract knowledge ensures that the interpretation is not specific to a particular machine to be monitored, but can be applied to the entire class of machines to which the criterion is applicable. By combining anomaly recognition and its physical interpretation, the device is ready to start and can therefore be cold-started even without previous collection of sensor data and data logging on the machine to be monitored.

[0015] The machine can be constructed from a single component or from a plurality of components separate from one another, which, for example, carry out a single production step or a plurality of production steps.

[0016] In a preferred embodiment, the machine learning unit includes an anomaly detection unit having a trained machine learning model that is adapted to the machine currently to be monitored based on sensor data recorded on the machine during a specified adaptation period during normal operation after being put into operation.

[0017] Therefore, the machine learning model trained for anomaly detection already provides optimized results for the machine to be monitored after this adaptation period.Since these sensor data are collected during normal operation of the machine, no manual annotation / labeling is required, i.e., assigning these sensor data to the operating mode of the machine.

[0018] In one advantageous embodiment, the machine learning model is adjusted at specified time intervals during machine operation.

[0019] The machine learning model can thus continuously adapt to changes that occur in the machine during operation, for example due to wear and tear or due to altered external conditions.

[0020] In a preferred embodiment, the abnormal result is output in the form of a binary value, which is determined based on the calculated abnormal value (also called abnormal score) and a limit value, wherein in particular the limit value is specified based on the empirical distribution of abnormal values ​​of non-abnormal data; or the abnormal result is output in the form of a weight, which indicates the prominence of the abnormality.

[0021] Therefore, the time when the exception occurs can be set flexibly.

[0022] In a preferred embodiment, the machine learning unit includes a data processing unit that receives raw data measured on the machine in chronological order and divides the raw data into processing parts over time and / or converts the data format of the raw data and / or merges multiple components of the raw data into a single parameter of sensor data and outputs it as sensor data.

[0023] Therefore, the data processing unit converts raw data received in different forms and at different time intervals into a uniform form, preferably as vector data.

[0024] In one advantageous embodiment, the data processing unit coordinates a subset of the raw data, whose input rate deviates from the output rate of the sensor data, to this output rate.

[0025] Thus, it is ensured that the signal rate synchronization of all sensor data is achieved, i.e., at any point in time when an output signal needs to be calculated from the sensor data, the values ​​of all input signals are present. The output signal of the data processing unit is a vector with elements associated with the anomaly detection unit. The output rate is a fixed specified value and corresponds to the rate at which anomaly detection is performed.

[0026] In an advantageous embodiment, the machine learning unit comprises an anomaly interpretation unit which receives the anomaly result and sensor data of the underlying processing part as input signals and outputs parameters of the sensor data sorted according to their relevance to the anomaly result as features.

[0027] By means of the anomaly interpretation unit, the judgment of the anomaly detection unit is presented in a traceable manner, preferably by means of an explainable machine learning method.

[0028] In an advantageous embodiment, the transfer unit receives sensor data of an underlying processing section and / or additionally sensor data of an adjacent processing section as input signals and determines therefrom the at least one indication for those features having the highest relevance.

[0029] The indication basically reflects the trend relative to the good state, ie the behavior of the characteristic, such as sharp rise, failure, oscillation, decline.

[0030] In an advantageous embodiment, the device further comprises an extraction unit which is designed to derive rules and instructions based on expert knowledge, textbook knowledge, a physical analysis description or a digital twin of the machine and to provide these rules and instructions to the interpretation unit before the device is put into operation.

[0031] Domain knowledge about the machines to be monitored and possible anomalies and their causes is preferably stored in an abstract manner as part of the interpretation unit. Storage in the form of abstract knowledge ensures that the interpretation is not specific to a particular machine to be monitored, but can be applied to those machines to which the rules apply, i.e., those belonging to a particular machine class.

[0032] In an advantageous embodiment, the interpretation unit is designed to receive and / or update further rules and / or instructions after commissioning.

[0033] Thus, new fault types or rules with corresponding indications can be incorporated and the interpretation unit can thus be adapted to the changed properties of the machine.

[0034] In an advantageous embodiment, the device comprises a user interface, which is designed to output abnormal results or rules and / or receive changes of abnormal results or changes of rules and / or new rules from a user.

[0035] This enables incorporating domain knowledge of experts in the extraction unit when updating the anomaly detection unit and criteria.

[0036] In an advantageous embodiment, when the abnormality result is negative, the interpretation unit outputs a confirmation of the normal state of the machine or outputs a warning prompt of a possible fault according to the instructions and the criteria.

[0037] Therefore, even in the case where the abnormality result is negative, a possible hint of an impending abnormality can be provided, or the abnormality result can be confirmed to be negative.

[0038] In an advantageous embodiment, the adjustment and / or updating of the machine learning model and / or the training of the anomaly interpretation unit and / or the extraction unit is performed in a server unit remote from the machine.

[0039] This reduces the processor load in the device by outsourcing computationally intensive processes to a remote server unit. By having a server unit with high computational power, the time taken for updating can also be reduced.

[0040] According to a second aspect, the invention relates to a method for monitoring the state of a machine, the method comprising the following steps:

[0041] -In the Machine Learning Unit (AD),

[0042] receiving sensor data collected on the machine during operation and including at least one parameter of the machine, and

[0043] determining an abnormal result and at least one feature that indicates the abnormal result from the sensor data and forwarding them to a transmission unit,

[0044] - determining, by the transmitting unit, at least one indication within the sensor data for the abnormal result according to the characteristic, and

[0045] forwarding the at least one indication to an interpretation unit (PI), and

[0046] - by means of the interpretation unit (PI), checking the abnormal result according to the at least one indication and a criterion containing a rule, the criterion comprising at least one assignment from the indication to a fault type of the machine, and

[0047] When the abnormality result is positive and the abnormality is confirmed, a control instruction for execution on the machine is output.

[0048] The method has the same advantages as the device.

[0049] A third aspect of the invention relates to a computer program product comprising a non-volatile computer-readable medium directly loadable into the memory of a digital computer, the non-volatile computer-readable medium comprising program code portions which, when executed by the digital computer, cause the digital computer to perform the steps of the method.

[0050] Unless otherwise indicated in the subsequent description, the terms "receiving", "determining", "transferring", "checking" and the like preferably relate to operations and / or processes and / or processing steps that change and / or generate data and / or transform these data into other data, wherein these data can be presented or exist in particular as physical quantities, for example as electrical impulses. The device and the units contained therein, such as a machine learning unit or a transfer unit, etc., can include one or more processors and are designed to perform the above-mentioned operations, processes or processing steps.

[0051] The processor may in particular be a central processing unit (CPU), a microprocessor or a microcontroller, such as an application specific integrated circuit or a digital signal processor, etc., which may be combined with a storage unit for storing program instructions.

[0052] A computer program product, such as a computer program device, can be provided or made available, for example, as a storage medium, such as a memory card, USB stick, CD-ROM, DVD, or also in the form of a downloadable file from a server in a network. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Exemplary embodiments of the method according to the invention and of the device according to the invention are shown by way of example in the drawings and are explained in more detail in the following description.

[0054] Figure 1 A schematic diagram shows a first embodiment of the device according to the invention in an application environment;

[0055] Figure 2 A detailed view of an exemplary embodiment of a device according to the invention with a functional unit is shown in a schematic diagram;

[0056] Figure 3 An embodiment of the method according to the invention is shown as a flow chart;

[0057] Figure 4 An exemplary embodiment of collected sensor data and the processing of these sensor data by the device according to the invention is shown schematically;

[0058] Figure 5An embodiment of an initial or repeated training of a functional unit of a device according to the invention is shown schematically; and

[0059] Figure 6 A second exemplary embodiment of the device according to the invention with a remote functional unit is shown in a schematic diagram.

[0060] Parts which correspond to one another are provided with the same reference symbols in all the figures. DETAILED DESCRIPTION

[0061] Figure 1 An application scenario is shown in which the state of a machine 10 is monitored by a monitoring device 20 according to the invention.

[0062] The machine 10 comprises, for example, one or more machine components of an industrial system, such as a press, a milling machine, a 3D printer, etc. in a production system, or a pump, a grinder, etc. in a conveying or distribution system. At least one sensor, usually a plurality of sensors, is arranged at or in the environment of the machine 10, which sensors detect, in particular, physical parameters of the machine 10 and output sensor signals with sensor data.

[0063] The monitoring device 20, also referred to hereinafter as the device 20, is modularly constructed and includes a machine learning unit 21, in which at least one machine learning model is implemented for anomaly recognition and for other analysis functions. The monitoring device 20 also includes an interpretation unit 23, in which domain knowledge including possible anomalies and possible fault causes is stored in an accessible and abstract manner. The device 20 includes a transmission unit 22, which processes the results and information from the machine learning unit 21 for physical interpretation by the interpretation unit 23.

[0064] The device 20 comprises an input / output interface 24, which provides a user interface with a user 25, such as a domain expert or a machine operator, for inputting and outputting data, in particular for outputting abnormal results or rules and / or for inputting changes to abnormal results or rules and / or for inputting new rules. Via the user interface, the control instructions determined by the interpretation unit 23 can be output, for example, to the machine operator for execution on the machine. In one embodiment, the input / output unit 24 comprises a data interface via which the control instructions are directly transmitted to the machine.

[0065] in accordance with Figure 22 to explain the method steps performed by the device 20. In a first method step S1, the machine learning unit 21 receives sensor data of the machine 10, which are collected while the machine 10 is in operation. In this case, the sensors contain data about at least one parameter, usually a plurality of different parameters, of the machine 10. In addition to physical parameters such as temperature, pressure, etc., the sensor data may also include information about the filling level of production resources or production materials of the machine 10 or images from the operation of the machine 10.

[0066] In the next step S2 , the machine learning unit determines an abnormal result and at least one feature proving the abnormal result and indicating the presence of an abnormality based on the sensor data, and forwards it to the transmission unit 22 .

[0067] In the transmission unit 22, for the abnormal result, at least one indication within the sensor data is determined according to the transmitted characteristics, see S3, and the indication is forwarded to the interpretation unit 23. At the interpretation unit, the abnormal result is checked according to the at least one indication and a criterion containing rules, see step S5, the criterion including at least one assignment from the indication to the fault type of the machine. In the case of a positive abnormal result indicating an abnormal state of the machine, a control instruction is output according to the determined fault type for execution on the machine 10, see step S6. In the case of a negative abnormal result, the interpretation unit outputs a confirmation of the normal state of the machine or an alarm prompt of a possible fault according to the set of descriptive indications and criteria.

[0068] Therefore, the machine learning unit not only provides an abnormal result, which explains whether the monitored machine is in normal operation (hereinafter also referred to as a good state) or in a bad state that deviates from the normal operation, but also provides the features that caused the abnormal result, such as a specific combination of parameters of these sensor data. The transmission unit translates these features into the "language" of the interpretation unit. The indication outputs, for example, the change process of the parameters A and B contained therein within the considered time period for the feature, such as "a sudden increase in parameter A" and "the average value of parameter B". Now, based on these physically interpretable indications and criteria, control instructions can be output. The control instructions are based on the fault type of the machine, and therefore output targeted and correct operating suggestions to the machine operator or control instructions to the control unit of the machine. Therefore, the results from the data-based machine learning unit are translated into physical phenomena by the transmission unit and mapped to the fault state and its elimination by the interpretation unit.

[0069] Now, according to Figure 3To elaborate in more detail the various units of the monitoring device 20 and the functional units of the monitoring device. The machine learning unit 21 includes a data processing unit PP, an anomaly detection unit AD and an anomaly interpretation unit AE. The transfer unit 22 includes a transfer function TR. The interpretation unit 23 includes an extraction unit RE and an interpretation function PI. The input / output interface 24 includes the user interface UI and the data interface MMI already mentioned.

[0070] The data processing unit PP receives raw data measured on the machine 10 in chronological order, and divides the raw data into processing parts over time, and converts the data format of the raw data when necessary, merges multiple components of the raw data into a single parameter, and outputs it to the machine learning unit.

[0071] The data processing unit PP receives sensor data incoming as raw data via the data interface MMI and converts these sensor data into vector input signals for the anomaly detection unit AD. That is, the sensor data points exist as data vectors. The raw data received in the data processing unit PP are signals from at least one, usually multiple sensors. These sensors, for example, perform measurements periodically and provide one-dimensional or multi-dimensional measurement points as sensor data for each measurement time point, such as images, spectra, hyperspectral images or scalars such as current or voltage. The measurement rates of the sensors may be different, and the sensor data are not necessarily generated synchronously. Regardless of the modality and dimensionality of the input data, the size of the sensor data does not change during the operation of the machine 10. For example, the resolution of the image or spectrum remains constant during operation. In addition, the data processing unit PP divides the raw data into processing parts over time according to the change process of the raw data values ​​or sensor data values, for example over time.

[0072] If necessary, the data processing unit PP coordinates the subset of raw data whose input rate deviates from the output rate to this output rate. For raw data whose measurement rate or input rate differs from the output rate, an adjustment of the sampling rate is performed. This can be achieved, for example, by means of linear interpolation, returning the last value, multi-rate filtering (e.g. (low-pass) filtering and change of the sampling rate), decomposing the input signal of the sensor data into time windows and averaging. In this case, the sensor data received between the last output time point and the current output time point are averaged. Further calculations are performed based on this average value.

[0073] In each of the above cases, the average is taken over time. Therefore, this can also be applied to multidimensional sensor signals, such as spectra or images. Thereby, the signal rate synchronization of all sensor data is ensured, that is, at any point in time when the output signal is to be calculated, the values ​​of all input signals are present. The output signal of the data processing unit PP is a data vector with elements associated with the anomaly detection unit AD. The output rate is a specified value and corresponds to the rate at which anomaly detection is performed. In addition, the processing part over time is determined based on the change process of the sensor data values ​​over time. The processing part over time is the following time period, in which each of the different parameters of these sensors has the same change process. For example, when the change process of at least one parameter over time changes, a new processing part over time is created. Therefore, these processing parts over time can have different time lengths.

[0074] The input signal of the anomaly detection unit AD is a data vector. The data vector has a length that is specified and fixed for a specific application, here the monitoring of the machine 10. The components of the input vector can be formed in different ways from unprocessed sensor data. This can be the output data of the data processing unit PP, or else it can be the raw sensor data received directly from the machine 10. The output signal of the anomaly detection unit AD is the anomaly result. The anomaly result is a binary value, which indicates whether an anomaly has been identified. As soon as there is a new input vector, a new value is generated, i.e. the input and output have the same update rate.

[0075] The function of the anomaly detection unit AD is performed with the aid of machine learning. For this purpose, there is a sufficient number of input vector samples, which describe the normal operation of the machine 10. When the monitoring device 20 is cold started without pre-training the machine learning unit 21 for the machine 10, a small first basic amount of sensor data or input data is sufficient to form an initial data set. From this initial data set, characteristics in these sensor data are learned, which distinguish non-abnormal situations (i.e. good states) from abnormal situations (i.e. bad states).

[0076] The anomaly detection unit AD is preferably designed as an unsupervised machine learning model, in particular as an autoencoder, or as an isolation forest, a one-class support vector machine, or as an outlier detection method. The anomaly result is output in the form of a binary value, which is determined according to the calculated outlier and a limit value, wherein in particular the limit value is specified according to the empirical distribution of the outliers of the non-abnormal data. Alternatively, the anomaly result is output in the form of a weight, which describes the prominence of the anomaly.

[0077] The anomaly value (also called anomaly score) output by the anomaly detection unit AD of the machine learning model is converted into a binary parameter by means of a limit value (also called a threshold) or is transferred as a "weight" that describes the "prominence" of the anomaly. In another embodiment, the threshold is specified based on a specified p-value, i.e., an exceedance probability or importance value, and an empirical distribution of anomaly values ​​of non-anomalous data.

[0078] In an initial step, the anomaly explanation unit AE needs to access the trained machine learning model of the anomaly detection unit AD and the training data used for training. The anomaly explanation unit AE receives the abnormal results and the sensor data of the underlying processing part as input signals. The anomaly explanation unit AE includes an explainable artificial intelligence method, such as a machine learning model constructed in a tree structure, using which the judgment of the anomaly detection unit AD can be presented in a traceable manner. The output signal of the anomaly explanation unit AE is a feature sorted according to the correlation with the judgment of the anomaly detection unit AD.

[0079] Figure 4An example of an output signal of the anomaly interpretation unit AE is shown in terms of a temporal course of sensor data of the machine 10. The sensor data measured on the machine 10 include: a parameter P1, such as a measured force; a parameter P2, such as a pressure; and a parameter P3, such as a temperature of the machine. Based on the sensor data P1, features M1 and M2 are calculated, such as a force averaged over a specified time period and a standard deviation of the force. Based on the parameter value P2, the average pressure and the standard deviation of the pressure are determined in the same way as features M3 and M4. The processing portion over time is determined by the data processing unit PP. For the processing portion t1, the anomaly interpretation unit AE outputs an output signal AER1. The order in which the features M1, ..., M5 are output in the output signal AER1 illustrates the relevance of these features when the anomaly result is determined by the anomaly detection unit AD. In the output signal AER1, feature M1 has the highest relevance. The relevance of the subsequent features M2, M3, M4, M5 decreases with respect to the respective previous feature. In the same way, output signals AER2 and AER3 are output for the processing portions t2, t3 over time. These processing sections over time can have different lengths, ie, temporal lengths. In processing section t2, feature M2 has the greatest impact on the abnormal results, while in processing section t3, feature M5 has the highest relevance.

[0080] In one embodiment variant, these time-dependent processing sections have a constant length. In another embodiment variant, more than one output signal can be output for a time-dependent processing section. This is the case in particular when the correlation difference between the two output signals is small. Thus, in Figure 4 For a single processing portion t4 over time comprising the processing portions t2 and t3 over time, two output signals may be output, here output signals AER2 and AER3.

[0081] The transfer unit 21 and the transfer function TR therein receive as input signals the features "interpreted" by the anomaly interpretation unit AE, which caused the judgment as "abnormal" in the anomaly detection unit AD. The transfer function TR additionally receives sensor data of the underlying processing part, and optionally additionally receives sensor data of the adjacent processing part as input signals. The transfer function TR translates this information into a "physically interpretable language", that is, into at least one indication, usually a set of indications, which describe one of the rules generated from the extraction unit RE. The indication basically reflects the trend relative to the good state, that is, the behavior of the feature, such as a sharp rise, failure, oscillation, decline, etc.

[0082] The output signal of the transmission unit 21 is the set of descriptive indications, which explain why the anomaly detection unit AD obtained an "abnormal" result. Optionally, the transmission unit 22 forwards the output signal of the anomaly interpretation unit AE, i.e., the ranking of the features according to their contribution to the current anomaly, to the interpretation unit 23.

[0083] The interpretation function PI of the interpretation unit 23 receives at least one indication from the transmission unit 22 as an input signal. The interpretation function PI checks the abnormal result based on the at least one indication and a rule-containing criterion, which includes at least one assignment from the indication to the fault type of the machine. The interpretation function PI then returns either a conclusion that it is actually a good state, such as one of a plurality of permitted operating states, or a conclusion that it is actually an abnormal state as an output signal. The control instructions stored in the interpretation function PI, such as operating suggestions, are fed back to the machine operator via the user interface UI of the input / output unit 24, or are returned to the machine 10 as an automatic response via the data interface MMI.

[0084] Via the extraction unit RE, the physics-based fault description is forwarded to the inspection device. An input signal of the extraction unit RE is received. In one embodiment, the extraction unit RE derives rules and indications based on expert knowledge, textbook knowledge, a physical analysis description or information of a digital twin of the machine. The "output signal" of the extraction unit RE is a list of rules stored in the interpretation function PI.

[0085] The monitoring device 20 can be readjusted at any time and in a simple manner if necessary. The monitoring device 20 itself provides an indication of the need for readjustment. The anomaly detection unit AD identifies, for example, a plurality of anomalies, but the interpretation function PI cannot assign any of them to a known rule. If the verification on the machine 10 results in a new type of fault or a new manifestation of a known fault, the adjustment / recreation of the rules is indicated in the extraction unit RE. The interpretation unit 23 is designed to receive and / or update further rules and / or instructions after commissioning.

[0086] If the verification on the machine 10 concludes that the operation is fault-free, but there are potential environmental changes, then in the anomaly detection unit AD, an adjustment of the machine model is indicated, for example by retraining. In one embodiment, the machine learning model is adjusted at specified time intervals during the operation of the machine.

[0087] The monitoring device with the described construction can be set up and operated for monitoring different machines. By way of example, the setting, introduction and operation phases of the monitoring device 20 for monitoring a hydraulic press are described. The sensors on the machine respectively determine the pressure, the force with a collection frequency of 1 Hz, and the press temperature and the ambient temperature of the press with a collection frequency of 1 / 60 Hz.

[0088] During the setup phase of the monitoring device 20 for monitoring the press, the following operations are performed, see Figure 5 :

[0089] The data processing device PP is configured to: receive sensor data from the above-mentioned sensors incoming at different sampling rates, and determine therefrom, for example, the mean value and variance for pressure and force respectively within a time window of one minute, as well as the press temperature in order to reduce the ambient temperature, and output it as a feature at a uniform output rate of 1 / 10 Hz.

[0090] The input format of the anomaly detection unit AD is adapted to the output format and output rate of the data processing unit PP. The anomaly detection unit AD comprises training and inference variants AD_T, AD_I of the machine learning model. The training variant AD_T is trained to the normal state of the press using sensor data / training data TD, which are collected during the normal operation of the press, and outputs anomaly results. The trained model of the training variant AD_T is output and used as an inference variant of AD_I. The inference variant receives the processed data as input signal, and outputs anomaly results as a binary value (true / false) for each input value.

[0091] The anomaly interpretation unit AE comprises training and inference variants of the machine learning models AE_T, AE_I. The training variant AE_T is trained based on the trained learning model of the anomaly detection unit AD and preferably all sensor data available up to the point in time when the anomaly interpretation unit AE is trained. The training variant AE_T identifies the importance of each combination of these features for a given instance of an anomaly result. The inference variant AE_I adopts the model of the training variant AE_T and is a function that, given a data vector and an anomaly result as input signals, outputs a list of feature combinations, sorted according to their importance for identifying the data vector as an anomaly.

[0092] The transfer function TR translates the output features of the feature combination into "physical" expressions, such as "normal, too high, too low, rising, stable, falling", so that these expressions can be compared with the expressions in the rules. To this end, the transfer function TR uses the importance of the features to the data vector at the time point tx considered, as well as the training data in a certain time period before the time point tx considered, as input signals. All features that are considered irrelevant are classified as "normal". For all relevant features, the state and trend are calculated using the training data and historical data as reference. For example, the temperature trend is calculated by considering the temperature in the previous hour. The average force is determined based on the average force value in the training data.

[0093] The extraction unit RE is populated with rules based on information from experts in the field of presses, which have qualitative characteristics of typical fault behaviors of presses. Information such as "cooling fault leads to continuous temperature increase and pressure drop of the press" is transformed into a rule for "cooling fault" in the form of "temperature increase trend, average pressure below normal value".

[0094] The explanation function PI is configured to call the rules from the extraction unit. The explanation function PI defines a matching function that determines the distance between the expected fault behavior and the current observation, for example by means of a distance function. The input signals of the explanation function PI are the current physical state of the press, i.e. the indication, and the expected physical state for various fault types. The output signals are fault candidate signals sorted according to importance.

[0095] After this setup phase, in an introduction phase, i.e. during an adaptation period, the units and functions of the monitoring device 10 can be put into operation in a state in which the monitoring device has not yet been trained with respect to the press. In this introduction phase, also called cold start, the data processing unit PP is already active. The monitoring device 20 can be used for practical applications without further calibration. The prerequisite is that the pump is in good condition. The monitoring device 20 first collects data without further evaluation and within a specified time period that is either fixed or defined by the user and for the first time receives a "start (re)training" command. If suspected fault situations are already present, the assignment of rules and markings / classifications for these situations is requested from the expert / user via the user interface.

[0096] The training variant AD_T of the anomaly detection unit is trained using these data and is passed as ADt to the inference variant AD_I and the training variant AE_T of the anomaly interpretation unit, see Figure 5 . Now, the anomaly detection unit AE_I is active and evaluates the sensor data collected at the pump. The training variant AE_T of the anomaly interpretation unit is trained based on the training variant AD_T of the anomaly detection unit and is handed over to the inference variant AE_I, see Figure 5 AEt in.

[0097] In a subsequent operating phase, the monitoring device processes the sensor data collected at the pump during operation of the pump as described and outputs control instructions. If the machine operator has a different assessment of the state assessed by the monitoring device 20, the machine operator can input a changed classification of the data point via the user interface, add it to the training data as further training data, and optionally initiate a retraining of the anomaly detection unit AD. As already described for the introduction phase, see Figure 5This is carried out as described. If the fault behavior of the pump changes as this operating phase progresses, one or more rules can be updated, deleted or added again by the extraction unit RE.

[0098] The units of the monitoring system can be implemented in a single physical unit. However, the units can also be arranged in different physical units, for example in a system unit 40 physically close to the monitored machine and a remote server unit 30, for example in the cloud, see Figure 6 . Here, the training variants and reasoning variants AD_T, AD_I of the anomaly detection unit and the training variants and reasoning variants AE_T, AE_I of the anomaly interpretation unit can be operated as independent components. It is particularly advantageous that the training variant AD_T of the anomaly detection unit, the training variant AE_T of the anomaly interpretation unit and the extraction unit RE are arranged and run in the remote server unit 40. Thus, computationally intensive tasks that are not relevant to the current assessment of the machine state can be outsourced. The units required for evaluating the sensor data are arranged in the system unit 40, in particular the data processing unit PP, the reasoning variants AD_I, AE_I of the anomaly detection unit and the anomaly interpretation unit, the transfer function TR and the interpretation function PI.

[0099] The monitoring device can be put into operation without complicated data collection and can be continuously adapted. By simple interactions that are to be realized quickly, the monitoring device can be continuously informed whether it is necessary to change the monitoring device in order to adapt. Once the monitoring device is put into operation on a machine, it can be used to monitor other machines of the same type without modification. Only the anomaly detection unit needs to be retrained or adjusted. If the machine type changes, the rules can be changed with the help of the extraction unit RE.

[0100] All method steps can be implemented by corresponding devices suitable for performing the corresponding method steps. All functions that can be performed by these subject features can be method steps of the method. All described and / or drawn features can be advantageously combined with each other within the scope of the present invention. The present invention is not limited to the described embodiments.

Claims

1. A monitoring device for monitoring the status of a machine (10), the monitoring device comprising (20): a machine learning unit (21), comprising an anomaly detection unit (AD) with a trained machine learning model, the machine learning unit being designed to: receive sensor data collected on the machine (10) during operation, comprising at least one parameter of the machine (10), and to determine an abnormal result and at least one parameter of the sensor data ranked according to relevance to the abnormal result as a feature proving the abnormal result, and to forward it to a transmission unit (22); the transmission unit (22) being designed to determine at least one indication for the abnormal result within the sensor data as a function of at least one of the features and to forward the at least one indication to an interpretation unit (23); as well as - the interpretation unit (23) is designed to check the abnormal result according to the at least one indication and a rule-containing criterion and, in the case of a positive abnormal result, to output a control instruction for execution on the machine, the criterion comprising at least one assignment from the indication to a fault type of the machine (10), Therein, the transfer unit (22) receives sensor data of an underlying processing part and / or additionally sensor data of an adjacent processing part as input signals and determines therefrom the at least one indication for those features having the highest relevance.

2. The device according to claim 1, wherein: The trained machine learning model is adapted to the machine (10) currently to be monitored based on sensor data recorded during a specified adaptation period during normal operation on the machine (10) after commissioning.

3. The device according to claim 2, wherein: The machine learning model is adjusted at specified time intervals during operation of the machine (10).

4. A device according to any one of the preceding claims, wherein: The abnormal result is output in the form of a binary value, which is determined based on the calculated abnormal value and the limit value, wherein in particular the limit value is specified based on the empirical distribution of abnormal values ​​of non-abnormal data; or the abnormal result is output in the form of a weight, which describes the prominence of the abnormality.

5. The device according to any one of the preceding claims, wherein: The machine learning unit includes a data processing unit (PP), which receives raw data measured on the machine (10) in chronological order, and divides the raw data into processing parts over time and / or converts the data format of the raw data and / or merges multiple components of the raw data into a single parameter and outputs it to the machine learning unit (21).

6. The device according to claim 5, wherein: The data processing unit (PP) coordinates a subset of raw data whose input rate deviates from the output rate of the sensor data with the output rate.

7. The device according to any one of claims 5 or 6, wherein: The machine learning unit (21) comprises an anomaly interpretation unit (AE), which receives the anomaly result and sensor data of at least one underlying processing part as input signals, and outputs parameters of the sensor data sorted according to relevance to the anomaly result as features to the transmission unit (22).

8. A device according to any one of the preceding claims, wherein: The interpretation unit (23) comprises an extraction unit (RE) which is designed to derive the rules and instructions based on expert knowledge, textbook knowledge, a physical analysis description or a digital twin of the machine and to provide the rules and instructions to the interpretation function (PI) before the device (20) is put into operation.

9. The device according to any one of the preceding claims, wherein: The interpretation unit (23) is designed to receive and / or update further rules and / or instructions after commissioning.

10. The device according to any one of the preceding claims, comprising a user interface (UI), wherein the user interface is designed to: output abnormal results or rules, and / or receive changes to the abnormal results or changes to the rules and / or new rules from a user.

11. The device according to any one of the preceding claims, wherein: In the case where the abnormal result is negative, the interpretation unit (23) outputs a confirmation of the normal state of the machine (10) or outputs an alarm prompt of a possible fault according to the set of descriptive instructions and the criteria.

12. A device according to any one of the preceding claims, wherein: In a remote server unit (40), the adjustment and / or updating of the machine learning model and / or the training of the anomaly interpretation unit (AE) is performed.

13. A method for monitoring the status of a machine (10), the method comprising the steps of: In a machine learning unit (21) comprising an anomaly detection unit (AD) with a trained machine learning model, - receiving (S1) sensor data collected on the machine (10) during operation and comprising at least one parameter of the machine (10), - determining (S2) an abnormal result and at least one parameter of the sensor data sorted according to relevance to the abnormal result as a feature proving the abnormal result from the sensor data, and forwarding it to a transmission unit (22), - determining (S3) by means of the transmitting unit (22) at least one indication within the sensor data for the abnormal result according to the characteristic, and - forwarding (S4) the at least one indication to an interpretation unit (23), and - checking (S5) the abnormal result by means of the interpretation unit (23) according to the at least one indication and a criterion containing rules, the criterion comprising at least one assignment from the indication to a fault type of the machine (10), and - in the case where the abnormality result is positive, confirming the abnormality, outputting (S6) a control instruction for execution on the machine (10) depending on the determined fault type, in, The transfer unit (22) receives sensor data of the underlying processing section and / or additionally sensor data of an adjacent processing section as input signals and determines therefrom the at least one indication for those features having the highest correlation.

14. A computer program product comprising a non-volatile computer-readable medium directly loadable into the memory of a digital computer, the non-volatile computer-readable medium comprising program code portions which, when executed by the digital computer, cause the digital computer to perform the steps of the method according to claim 13.

Images