Machine learning based condition monitoring with adjusted first physical quantity

Abstract

Embodiments of the present disclosure relate to machine learning based condition monitoring with an adjusted first physical quantity. A method of adjusting data used to detect a fault condition of a machine, comprising: estimating, in a first machine learning model (30), required output data (OD) based on input data, the input data comprising at least one measured first physical quantity (PQ1M) associated with the machine, and the output data (OD) comprises at least one estimated second physical quantity (PQ2E) of the machine; processing at least some of the input data for acquiring an adjusted first physical quantity (PQ1E), wherein an external influence of the adjusted first physical quantity (PQ1E) has been limited; and applying the adjusted first physical quantity (PQ1E) together with input data not including the measured physical quantity (RI) in a first machine learning model (30) for obtaining modified output data (OD *) for use in detecting a fault condition of the machine, the modified output data (OD *) including at least one adjusted estimated second physical quantity (PQ2E), the external influence has been limited in the at least one adjusted estimated second physical quantity (PQ2E).

Description

Technical Field

[0001] The technologies disclosed herein generally relate to the field of machine monitoring, and in particular to a method, data adjustment device, computer program, and computer program product for adjusting data used to detect machine malfunctions. Background Technology

[0002] Condition monitoring of machines (such as electric motors with or without motor drives) is very important because it can identify and resolve faults before they become problems.

[0003] For example, US2021 / 0341896 discloses an industrial motor drive with condition monitoring. This document recommends equipping the motor drive with an embedded analytics engine, rather than performing condition monitoring externally to the motor drive.

[0004] Machine learning can be used for condition monitoring. It can be used to estimate output data based on measured input data.

[0005] However, there is a problem: the physical quantities measured as part of the input data may be unreliable. For example, if there is a fan problem in the machine, the ambient temperature, typically measured at the air intake, may be affected. This could also affect the estimated output data.

[0006] Therefore, there may be a problem that the physical quantities used as input in machine learning models for condition monitoring purposes are unreliable.

[0007] The machine may include several components. Many of these components may also be equipped with temperature sensors, for example, embedded in windings, bearings, and frames. In electric drives, ambient temperature and power module temperature are also monitored. Temperature signals from these sensors (such as wiring-free smart sensors) can be used to issue alarms when set thresholds are exceeded. However, if a component reaches such a threshold, either the failure is severe, causing a significant temperature rise, or the component is already operating close to its maximum capacity before the temperature rises. In either case, the component may experience adverse operating conditions under these circumstances, thus shortening its lifespan.

[0008] Temperature signals are commonly used to detect malfunctions in machines. In most cases, they serve as a safety measure: when a temperature threshold is violated, an alarm is triggered, and the machine stops operating. Therefore, effective monitoring is only possible when the machine's performance is at its higher limits. If conditions are far below this point, the temperature rise needs to be very significant to reach the threshold. This, in turn, can adversely affect the machine's lifespan.

[0009] To accurately estimate the temperature, a known temperature must be referenced. The commonly used reference temperature is either the coolant temperature or the ambient temperature. However, in some cases, a fault can affect the reference temperature, thus weakening evidence of a fault.

[0010] All aspects of machine monitoring need to be addressed and improved. Summary of the Invention

[0011] This invention relates to the problem of unreliable first physical quantity as input in machine learning models used for monitoring the condition of machines.

[0012] Therefore, one object of the present invention is to improve the reliability of the first physical quantity used as input in a machine learning model for monitoring the condition of a machine.

[0013] According to the first aspect, this objective is achieved by a method for adjusting data used to detect machine malfunctions, the method comprising:

[0014] - In a first machine learning model, the desired output data is estimated based on input data, which includes at least one measured first physical quantity associated with the machine, and the output data includes at least one estimated second physical quantity of the machine;

[0015] - Process at least some of the input data to obtain an adjusted first physical quantity, wherein the external influences of the adjusted first physical quantity have been constrained; and

[0016] - An adjusted first physical quantity, together with input data excluding the measured first physical quantity, is applied to a first machine learning model to obtain modified output data for use in detecting fault conditions of a machine. The modified output data includes at least one adjusted estimated second physical quantity in which external influences have been constrained.

[0017] According to the second aspect, this objective is achieved by a data adjustment device for adjusting data used to detect machine malfunctions, the data adjustment device including a processor operable to:

[0018] - In a first machine learning model, desired output data is estimated based on input data, which includes at least one measured first physical quantity associated with the machine, and the output data includes at least one estimated second physical quantity from the machine.

[0019] - Process at least some of the input data to obtain an adjusted first physical quantity, wherein the external influences of the adjusted first physical quantity have been constrained; and

[0020] - An adjusted first physical quantity, together with input data excluding the measured first physical quantity, is applied to a first machine learning model to obtain modified output data for use in detecting fault conditions of a machine. The modified output data includes at least one adjusted estimated second physical quantity in which external influences have been constrained.

[0021] According to the third aspect, this objective is achieved by a computer program for adjusting data used to detect machine malfunctions, the computer program comprising computer program code that, when executed by the processor of the data adjustment device, causes the data adjustment device to:

[0022] - In a first machine learning model, the desired output data is estimated based on input data, which includes at least one measured first physical quantity associated with the machine, and the output data includes at least one estimated second physical quantity of the machine;

[0023] - Process at least some of the input data to obtain an adjusted first physical quantity, wherein the external influences of the adjusted first physical quantity have been constrained; and

[0024] - An adjusted first physical quantity, together with input data excluding the measured first physical quantity, is applied to a first machine learning model to obtain modified output data for use in detecting fault conditions of a machine. The modified output data includes at least one adjusted estimated second physical quantity in which external influences have been constrained.

[0025] According to the fourth aspect, this objective is achieved by a computer program product for adjusting data used to detect the fault condition of a machine, the computer program product including a data carrier having the computer program according to the third aspect.

[0026] The process may include estimating a first physical quantity based on input data excluding the measured first physical quantity, wherein the estimation of the first physical quantity may be performed in a second machine learning model. Alternatively, or as an alternative, an analytical model of the first physical quantity may be used to estimate the first physical quantity.

[0027] The analytical model can be one in which the first physical quantity is expressed as a function of the machine’s operating condition and the environmental condition.

[0028] In the first machine learning model, estimating the desired output data based on the input data can be done in a first instance of the first machine learning model, and applying the adjusted first physical quantity to the first machine learning model can include applying the adjusted first physical quantity together with input data excluding the measured first physical quantity to the first instance of the first machine learning model. In this case, applying the adjusted first physical quantity to the first machine learning model can include replacing the measured first physical quantity with the adjusted first physical quantity in the first machine learning model (i.e., the first instance of the first machine learning model).

[0029] Alternatively, applying the adjusted first physical quantity to the first machine learning model may include: applying the estimated first physical quantity together with input data excluding the measured first physical quantity to a second instance of the first machine learning model.

[0030] According to a variation of the first aspect, the method further includes: investigating the measured first physical quantity in relation to the adjusted first physical quantity, and determining, based on the investigation, that the adjusted first physical quantity will be used to detect the fault condition of the machine.

[0031] According to the second aspect of the strain type, the data adjustment device can also be operated to investigate the first physical quantity measured against the adjusted first physical quantity, and based on the investigation, determine that the adjusted first physical quantity will be used to detect the fault condition of the machine.

[0032] In this scenario, the investigation may include comparing the difference between the measured first physical quantity and the adjusted first physical quantity with an input threshold, and if the input threshold is exceeded, determining that the adjusted first physical quantity will be used. The investigation may also include comparing the absolute value of the difference with the input threshold.

[0033] According to a variation of the first aspect, the method further includes: acquiring a measurement of a second physical quantity of the machine.

[0034] According to the second aspect of the strain type, the data adjustment device can also be operated to obtain the measurement value of the second physical quantity of the machine.

[0035] When acquiring a measurement of the second physical quantity, the method may further include: investigating the measured second physical quantity in light of an estimate of the second physical quantity based on the measured first physical quantity, and determining, based on the investigation, that the adjusted first physical quantity will be used to detect the machine's malfunction condition.

[0036] When the measurement value of the second physical quantity is acquired, the data adjustment device can also be operated to: investigate the measured second physical quantity based on the estimate of the second physical quantity based on the measured first physical quantity, and determine, based on the investigation, that the adjusted first physical quantity will be used to detect the fault condition of the machine.

[0037] In the above scenario, the investigation may include: comparing the difference between the measured second physical quantity and the estimated second physical quantity with a corresponding first output threshold, and determining that an adjusted first physical quantity will be used if the first output threshold is exceeded. The investigation may also include: comparing the absolute value of the difference with the first output threshold.

[0038] When acquiring a measurement of the second physical quantity, the method may further include: investigating the measured second physical quantity for an estimated second physical quantity output by a second instance of the first machine learning model, and generating an alarm based on the investigation.

[0039] When acquiring a measurement of the second physical quantity, the data adjustment device can also operate to: estimate the second physical quantity based on the output of a second instance of the first machine learning model, investigate the measured second physical quantity, and generate an alarm based on the investigation.

[0040] The investigation may include comparing the difference between the measured second physical quantity and the estimated second physical quantity with a corresponding second output threshold, and generating an alarm if the second output threshold is exceeded. The investigation may also include comparing the absolute value of the difference with the second output threshold.

[0041] The first physical quantity can be a physical quantity of the environment surrounding the machine, such as the ambient temperature of the machine. The second physical quantity can be a physical quantity of the same type as the first physical quantity. If the first physical quantity is temperature, then the second physical quantity can also be temperature. Alternatively, the second physical quantity can be another type that can be derived from the type of the first physical quantity. For example, it can be pressure.

[0042] The first physical quantity can also be of other types besides temperature, such as pressure.

[0043] The first machine learning model can be a thermal machine learning model of the machine, where the output data includes the machine's internal temperature. The second machine learning model can also be a thermal machine learning model of the machine.

[0044] Further objects, features, and advantages of the appended embodiments will become apparent from the following detailed disclosure, the appended dependent claims, and the accompanying drawings.

[0045] Generally, unless otherwise expressly defined herein, all terms used in the claims shall be interpreted according to their ordinary meaning in the art. Unless otherwise expressly stated, all references to “a / an / the element, device, component, apparatus, module, action, etc.” shall be explicitly interpreted as referring to at least one instance of the element, device, component, apparatus, module, action, etc. Unless otherwise expressly stated, the actions of any method disclosed herein need not be performed in the order disclosed. Attached Figure Description

[0046] The inventive concept will now be described by way of example with reference to the accompanying drawings.

[0047] Figure 1 This is a schematic diagram of a machine in the form of a motor connected to a motor driver, which drives a load.

[0048] Figure 2 An implementation of a data adjustment device for adjusting data used to detect machine malfunctions is schematically shown, the data adjustment device being implemented as a processor executing computer instructions in memory.

[0049] Figure 3 A computer program product having computer instructions for implementing data adjustment functionality of a data adjustment device is schematically shown.

[0050] Figure 4 A flowchart of a first embodiment of a method for adjusting data used to detect machine malfunctions is shown.

[0051] Figure 5 This schematically illustrates one implementation of the functionality of a data adjustment device.

[0052] Figure 6 It shows the relationship with Figure 5 The flowchart shows a second embodiment of a method for adjusting data used to detect machine malfunctions, corresponding to the implementation method described in the diagram.

[0053] Figure 7 This schematically illustrates another way of implementing the functionality of a data adjustment device.

[0054] Figure 8 It shows the relationship with Figure 7 The flowchart shows a second embodiment of a method for adjusting data used to detect machine malfunctions, corresponding to the implementation method described in the diagram.

[0055] Figure 9 Flowcharts of several additional method steps used in the second and third embodiments are shown.

[0056] Figure 10 and Figure 11 Two alternative implementations of data adjustment functionality are shown. Detailed Implementation

[0057] The inventive concept will now be described more fully with reference to the accompanying drawings, in which certain embodiments of the inventive concept are illustrated. However, the inventive concept can be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided by way of example only so that this disclosure will be thorough and complete, and will fully convey the scope of the inventive concept to those skilled in the art. Throughout the specification, the same numerals refer to the same elements.

[0058] Figure 1 An example of a machine MA 10 capable of performing condition monitoring is schematically shown. In this example, machine 10 includes a motor M12 connected to a motor driver MD 14. Motor 12 also drives a load L16. It should be noted that the motor driver can be omitted from the machine.

[0059] The machine 10 can be condition monitored, wherein the condition monitoring investigates one or more physical quantities in order to determine the health of at least a portion of the machine.

[0060] In such investigations, the use of machine learning is also of interest. For example, input data, including at least one measurable physical quantity associated with the machine, can be fed to a machine learning model to produce output data including one or more estimated physical quantities of the machine to be monitored. The estimated physical quantities can then be used to provide an indication of the health of at least a part of the machine.

[0061] However, one or more of the physical quantities measured in this way may be unreliable, possibly due to a malfunction inside or around the machine. If these physical quantities are unreliable, one or more of the estimated output quantities may also be incorrect. This can lead to errors in condition monitoring.

[0062] This disclosure addresses this issue in various aspects.

[0063] According to various aspects of this disclosure, one device for solving this problem is a data conditioning device. Figure 2 The diagram illustrates one implementation of the data adjustment device DAD 18. The data adjustment device 18 can be implemented as a processor 20 having an associated program memory 22, which includes a computer program having computer program code or computer instructions CI 24 for implementing data adjustment functions.

[0064] Computer programs can also be provided via computer program products, such as in the form of computer-readable storage media or data carriers (e.g., CD-ROMs or memory sticks), which carry computer program code that, when loaded into a processor, will implement data manipulation functions. Figure 3The diagram schematically illustrates a computer program product in the form of a CD ROM 26 having the aforementioned computer program code 16 including computer program code or computer instructions CI 24.

[0065] Now refer to Figure 4 The operation of the device is described based on the data from the first embodiment. Figure 4 Several method steps are shown in a method for adjusting data used to detect fault conditions of machine 10.

[0066] As previously mentioned, machine learning can be used to perform condition monitoring. For this reason, the data adjustment function of the data adjustment device 18 includes a first machine learning model that receives measured input data and estimates output data based on the measured input data, wherein the input data includes a first physical quantity PQ1 measured and associated with the machine 10. M The output data OD includes at least one estimated second physical quantity PQ2 from machine 10. E .

[0067] The first machine learning model could be a thermal machine learning model of the machine, where the output data includes the machine's internal temperature.

[0068] For example, the first physical quantity could be the ambient temperature of the machine or the rack temperature of the motor, and the second physical quantity could be other temperatures, such as rack temperature or power module temperature.

[0069] In addition to the primary physical quantity, the input data may also include driver and / or motor input data, such as motor speed and current, as well as the switching frequency, output current, and output power of the drive stage. The input data may also undergo post-processing, such as by calculating a moving average, determining the root mean square (RMS) value, and applying filtering (e.g., Butterworth filtering).

[0070] Therefore, the first physical quantity can be a physical quantity of the environment surrounding the machine, such as the ambient temperature of the machine. The second physical quantity can be a physical quantity of the same type as the first physical quantity. If the first physical quantity is temperature, the second physical quantity can also be temperature. Alternatively, the second physical quantity can be of other types that can be derived from the type of the first physical quantity. For example, it could be pressure.

[0071] The first physical quantity can also be of other types besides temperature, such as pressure.

[0072] According to the first embodiment, the operation of the data adjustment device includes: S100, estimating the required output data OD based on input data in a first machine learning model ML1, wherein the input data includes at least one first physical quantity PQ1 associated with machine 10. MSuch as being related to the environment surrounding the machine; and the output data OD includes at least one estimated second physical quantity PQ2 of machine 10. E Such as the temperature inside the machine. Input data can be measured in or at machine 10 and then transmitted to data adjustment device 18.

[0073] The data adjustment function further includes: S110, processing at least some of the input data to obtain an adjusted first physical quantity, wherein the external influence of the adjusted first physical quantity has been limited. This processing may involve estimating the first physical quantity in a second machine learning model based on input data excluding the measured first physical quantity. Therefore, the remaining input data can be used to estimate the first physical quantity in the second machine learning model. Thus, if the accuracy of the measured first physical quantity decreases, a more accurate first physical quantity than the one being measured can be obtained. As an alternative to using a second machine learning model, the data adjustment device may include a method for adjusting the measured first physical quantity PQ1. M Filtering and / or delay stages are performed. This allows for the removal of at least some inaccuracies in the measured first physical quantity, and / or delay of the measured first physical quantity PQ1. M The impact of inaccuracies. Therefore, the external influence on the first physical quantity can be limited by estimating the first physical quantity based on the remaining input data or by filtering and / or delaying the measured first physical quantity.

[0074] Finally, the data adjustment function includes: S120, adjusting the first physical quantity and the first physical quantity PQ1 excluding the measurement. M The input data is used together with the first machine learning model ML1 to obtain modified or adjusted output data OD* for use in detecting machine failure conditions. The modified output data OD* includes an adjusted estimate of the second physical quantity PQ2. E External influences have been limited.

[0075] The estimation of the desired output data OD based on the input data can be made in a first instance of the first machine learning model. Then, the application of the adjusted first physical quantity to the first machine learning model can also be made in a first instance of the first machine learning model. Alternatively, it can also be applied in a second instance of the first machine learning model.

[0076] Therefore, even if the first physical quantity is measured inaccurately, a more accurate estimate of the second physical quantity can be obtained. This can improve condition monitoring. This can be helpful in various situations, such as when deciding whether to perform maintenance.

[0077] When using a second machine learning model, it can also be a thermal machine learning model of the machine, where the output includes the machine's ambient temperature.

[0078] The machine learning model can be a trained machine learning model. Post-processed input data can be used for all machine learning models. The primary machine learning model can also be a machine learning model used to predict the temperature of the drive frame or power module and / or the target temperature of the motor (such as rotor fins, windings, DE bearings, and NDE bearing temperatures).

[0079] Although temperature is given as an example, it should be recognized that other physical quantities, such as pressure, can also be used. The first physical quantity could be, for example, the machine's supply pressure or inlet pressure, and the second physical quantity could be the machine's relevant internal pressure or coolant pressure.

[0080] Now refer to Figure 5 and Figure 6 The second embodiment is described, wherein Figure 5 This schematically illustrates one implementation of the functionality of a data adjustment device, and Figure 6 A flowchart of several steps in a second embodiment of a method for adjusting data used to detect machine malfunctions is shown.

[0081] like Figure 5 As can be seen, the data adjustment device 18 implements a first instance 30 of the first machine learning model ML1 and a second machine learning model ML2 32, wherein the second machine learning model 32 receives a first physical quantity PQ1 that is not measured. M The input data (i.e., the remaining input data RI) is used as input, and the first physical quantity PQ1 is estimated. E As output, the first instance 30 of the first machine learning model also receives input data excluding the measured first physical quantity, i.e., the remaining input data RI. It also receives the measured first physical quantity PQ1. M Or the estimated first physical quantity PQ1 E As input, a switch can be used to select which version of the first physical quantity to receive. The first instance 30 of the first machine learning model provides an estimate of the second physical quantity PQ2, including (unadjusted) values. E The unmodified output data OD, or a second physical quantity PQ2 with an adjusted estimate, is provided. E The modified output data OD*, wherein providing either the unmodified output data OD or the modified output data OD* depends on the version of the first physical quantity used in the first instance 30 of the first machine learning model. Therefore, for the estimated second physical quantity PQ2... E The adjustment is made by using the estimated first physical quantity PQ1 in the first instance 30 of the first machine learning model.E This is achieved through [the means].

[0082] In this embodiment, the operation may include: S200, estimating the desired output data OD based on the input data in the first instance 30 of the first machine learning model, wherein the input data includes a measured first physical quantity PQ1. M The output data includes an estimated second physical quantity, PQ2. E The operation also includes: S210, in the second machine learning model 32, based on the first physical quantity PQ1, which does not include measurement. M The first physical quantity PQ1 is estimated based on the input data (i.e., based on the remaining input data RI). E ; and S220, in the first instance 30 of the first machine learning model, using the estimated first physical quantity PQ1 E To replace the first physical quantity PQ1 being measured M Therefore, the output data is the modified output data OD*, which has an adjusted estimate of the second physical quantity PQ2. E This modification is based on using the estimated first physical quantity PQ1. E The changes are made based on the condition. Therefore, in condition monitoring, the process switches from using unmodified output data OD, which includes the (unadjusted) estimated second physical quantity, to using modified output data OD*, which has an adjusted estimate of the second physical quantity.

[0083] This has the following advantages: it provides a good (adjusted) estimate of the second physical quantity using a limited amount of processing, and the influence of errors in the measured first physical quantity is limited.

[0084] Now refer to Figure 7 and Figure 8 The third embodiment is described, wherein Figure 7 This schematically illustrates another implementation of the functionality of the data adjustment device 18, and Figure 8 A flowchart of several method steps is shown in a third embodiment of a method for adjusting data used to detect machine malfunctions.

[0085] like Figure 7 As can be seen, the first instance 30 of the first machine learning model ML1 receives the raw input data. Therefore, it receives the first physical quantity PQ1 to be measured. M And the remaining input data RI, and provide a second physical quantity PQ2 with an (unadjusted) estimate. E The unmodified output data OD is used as the output. In addition, there is a second machine learning model ML2 32, which receives input data (i.e., the remaining input data RI) excluding the measured first physical quantity as input and provides an estimated first physical quantity PQ1.E As output, the estimated first physical quantity PQ1 E The remaining input data RI is provided as input to the second instance 34 of the first machine learning model ML1. The second instance 34 of the first machine learning model provides modified output data OD*, which includes the adjusted estimated second physical quantity PQ2. E .

[0086] In this embodiment, the operation includes: S300, in the first instance 30 of the first machine learning model, estimating the desired output data OD based on the input data, wherein the input data includes a measured first physical quantity PQ1. M The unmodified output data OD includes the (unadjusted) estimated second physical quantity PQ2. E The operation also includes: S310, in the second machine learning model 32, estimating the first physical quantity PQ1 based on input data that does not include the measured first physical quantity RI. E ; and S320, which estimates the first physical quantity PQ1 E The first physical quantity PQ1, which is not included in the measurement M The input data, together with the first machine learning model, is applied to a second instance 34 to obtain a second physical quantity PQ2, including an adjusted estimate. E The modified output data OD*.

[0087] In this embodiment, there are two parallel processing paths for generating unmodified output data OD and modified output data OD* via two instances of the first machine learning model. The modified output data OD* is used in condition monitoring by selecting the output of the second instance 34 of the first machine learning model, rather than the output of the first instance 30. Therefore, this is a selection of the output signal used in condition monitoring, rather than the selection of the input signal.

[0088] This has the advantage of quickly obtaining a better, adjusted estimate of the second physical quantity. This improvement can be achieved because the estimate of the first physical quantity can be used to train a second instance of the first machine learning model.

[0089] As can be seen above, in all embodiments, the condition monitoring is changed from using a measured first physical quantity to using an adjusted or estimated first physical quantity.

[0090] Now refer to Figure 9 Describe how to implement this change for the second and third embodiments. Figure 9 Flowcharts of several other method steps are shown.

[0091] To determine when to use the first physical quantity PQ1 for measurement M Switch to using the estimated first physical quantity PQ1 E It is necessary to target the first physical quantity PQ1 for estimation. E The first physical quantity to be investigated and measured is PQ1 M The investigation may include: S400, determining the first physical quantity PQ1 to be measured. M With the estimated first physical quantity PQ1 E The difference between them can be referred to as the first difference. Then, in step S410, the first difference (more specifically, the absolute value of the first difference) can be compared with the input threshold TH. I Compare the differences. The absolute value of the first difference is below the input threshold TH. I In the case of (S420), the first physical quantity PQ1 is measured. M It continues to be used for condition monitoring, specifically the S480, which is used to detect machine malfunctions.

[0092] However, if the input threshold TH is exceeded I That is, the absolute value of the first difference is higher than the input threshold TH. I Then, the second physical quantity PQ2 measured by machine 10 is obtained. M S430. The measurement can be made in the machine and transmitted to the data adjustment device 18. Then, for the first physical quantity PQ1 measured, M The second physical quantity PQ2 is estimated by the first instance 30 of the first machine learning model. E The second physical quantity PQ2 being measured M An investigation is conducted. In this case, the investigation includes: S440, determining the second physical quantity PQ2 to be measured. M With the estimated second physical quantity PQ2 E The difference between them, which can be referred to as the second difference; and S450, comparing the measured second difference with the corresponding first output threshold TH O1 A comparison is made. This comparison could be the absolute value of the second difference versus the first output threshold TH. O1 The comparison. When the absolute value of the second difference is lower than the first output threshold TH. O1 In the case of (S460), the first physical quantity PQ1 is measured. M It continues to be used for condition monitoring, S480. However, when the first output threshold TH is exceeded... O1 In the case of (S460), that is, if the second difference is higher than the first output threshold TH O1 Then determine the first physical quantity PQ1 to be estimated. E It will be used for condition monitoring, that is, to detect the malfunction of the machine.

[0093] Therefore, the determination of the first physical quantity to be used for condition monitoring, i.e., detecting the fault condition of the machine, is based on the investigation of the first physical quantity and the investigation of the second physical quantity.

[0094] It should be recognized here that it is possible to investigate only the first physical quantity or the second physical quantity.

[0095] Optionally, the measured second physical quantity can be investigated based on the adjusted estimated second physical quantity provided by the second instance of the first machine learning model, and an alert can be generated based on the investigation.

[0096] The subsequent investigation may include determining the second physical quantity PQ2 to be measured. M With the adjusted estimate of the second physical quantity PQ2 E The difference between the two thresholds, referred to as the third difference, is compared to the corresponding second output threshold. This comparison can be the absolute value of the third difference versus the second output threshold. An alarm can be generated if the absolute value of the second difference is higher than the second output threshold.

[0097] As mentioned earlier, the estimated first physical quantity can be replaced by a filtered or delayed first physical quantity. Figure 10 and Figure 11 Variations of the second and third embodiments are illustrated schematically, wherein the second machine learning model has been replaced by a filtering / delay stage 36, which provides a delayed and / or filtered first physical quantity PQ1. D / F As the first physical quantity after adjustment. Furthermore, the filter / delay stage 36 only receives the measured first physical quantity PQ1. M It does not receive the remaining input data RI.

[0098] According to another variation, the first physical quantity can be estimated using an analytical model of the first physical quantity based on the remaining input data RI. This analytical model can replace or be provided as part of a second machine learning model.

[0099] The analytical model can be one in which the first physical quantity is expressed as a function of the machine’s operating conditions and environmental conditions, wherein the operating conditions may include information such as motor speed, motor current and output frequency, and the environmental conditions may include reference temperature and cooling method.

[0100] The different methods described above for replacing unmodified output data with modified output data can also be used in the above variations.

[0101] The inventive concept has been described above primarily with reference to several embodiments. However, those skilled in the art will readily understand that, within the scope of the inventive concept, other embodiments are also possible besides those disclosed above, as defined in the appended claims.

Claims

1. A method of adjusting data used for detecting a fault condition of a machine (10), the method comprising: In the first machine learning model (30), the required output data (OD) is estimated (S100; S200; S300) based on input data, said input data including at least one measured first physical quantity (PQ1) associated with the machine (10). M The output data (OD) includes at least one estimated second physical quantity (PQ2) of the machine (10). E ); processing (S110; S210; S310) at least some of the input data for obtaining an adjusted first physical quantity (PQ1 E ; PQ1 D / F ), wherein an external influence on the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) has been limited; and applying the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) to the first machine learning model (30; 34) together with the input data not including the measured first physical quantity (RI) for obtaining modified output data (OD*) for use in detecting a fault condition of the machine, the modified output data (OD*) comprising at least one adjusted estimated second physical quantity (PQ2 E ), in which at least one adjusted estimated second physical quantity PQ2 E the external influence has been limited.

2. The method of claim 1, wherein the processing comprises: estimating (S210; S310) the first physical quantity (PQ1 E ) based on the input data not including the measured first physical quantity (RI).

3. The method according to claim 2, wherein the estimation (S210; S310) of the first physical quantity (PQ1 E ) is performed in a second machine learning model (32).

4. The method according to claim 2 or 3, wherein the estimate of the first physical quantity (PQ1 E ) is made using an analytical model of the first physical quantity.

5. The method according to any one of the preceding claims, wherein the estimating (SI 00; S200; S300) is made in a first instance (30) of the first machine learning model, and applying the adjusted first physical quantity to the first machine learning model comprises: applying (S330) the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) to a second instance (34) of the first machine learning model together with the input data not including the measured first physical quantity.

6. The method of claim 5, further comprising: acquiring (S430) a measured value of the second physical quantity (PQ2 M ) of the machine (10); investigating the measured second physical quantity against the estimated second physical quantity output by the second instance of the first machine learning model; and generating an alert based on the investigation.

7. The method of any of the preceding claims, further comprising: acquiring (S430) a measured value of the second physical quantity (PQ2 M ) of the machine (10); investigating the measured second physical quantity (PQ2 M ) based on the estimate of the first physical quantity (PQ1 M ) for the second physical quantity (PQ2 E ); and determining (S470) based on the investigation that the adjusted first physical quantity (PQ1A; PQ1 E ) is to be used for detecting a fault condition of the machine.

8. The method of claim 7, wherein the investigation comprises: the measured second physical quantity (PQ2 M ) and the estimated second physical quantity (PQ2 E ) is compared (S450) with a corresponding output threshold (TH O1 ); and if the output threshold (TH O1 ) is exceeded, the determination that the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) is to be used is made (S460).

9. The method of any one of claims 1 to 4, 7, or 8, wherein applying the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) to the first machine learning model comprises: In the first machine learning model (30), the measured first physical quantity is replaced (S230) with the adjusted first physical quantity.

10. The method of any of the preceding claims, further comprising: For the adjusted first physical quantity (PQ1) E PQ1 D / F The first physical quantity to be measured (PQ1) was investigated. M ), and based on the survey, determine (S470) the adjusted first physical quantity (PQ1). E PQ1 D / F This will be used to detect malfunctions in the machine.

11. The method of claim 10, wherein the investigation comprises: comparing (S410) a difference between the measured first physical quantity (PQ1 M ) and the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) with an input threshold (TH I ); and If the input threshold (TH I ) is exceeded, the determination that the adjusted first physical quantity will be used to obtain the required output data is made (S420).

12. The method according to any one of the preceding claims, wherein the first physical quantity is a physical quantity of an environment surrounding the machine, such as an ambient temperature.

13. A data adjustment device (18) for adjusting data used for detecting a fault condition of a machine (10), the data adjustment device (18) comprising a processor (20) operable to: estimating required output data (OD) in a first machine learning model (30) based on input data, the input data comprising at least one measured first physical quantity (PQ1 M ) associated with the machine (10) and the output data comprising at least one estimated second physical quantity (PQ2 E ) of the machine (10); processing at least some of the input data for obtaining an adjusted first physical quantity (PQ1 E ; PQ1 D / F ), wherein an external influence on the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) has been limited; and The adjusted first physical quantity (PQ1) E PQ1 D / F The input data, excluding the first physical quantity (RI) measured, is applied to the first machine learning model (30; 34) to obtain modified output data (OD*) for use in detecting fault conditions of the machine (10), the modified output data (OD*) including at least one adjusted estimated second physical quantity (PQ2). E ), in the at least one adjusted estimated second physical quantity (PQ2) E The external influences described in the document have been limited.

14. A computer program for adjusting data used for detecting a fault condition of a machine (10), the computer program comprising computer program code (24) which, when run by a processor (20) of a data adjustment device (18), causes the data adjustment device (18) to: estimating required output data (OD) in a first machine learning model (30) based on input data, said input data comprising at least one measured first physical quantity (PQ1 M ) associated with said machine, and said output data comprising at least one estimated second physical quantity (PQ2 E ) of said machine (10); processing at least some of the input data for obtaining an adjusted first physical quantity (PQ1 E ; PQ1 D / F ) wherein an external influence on the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) has been limited; and applying the adjusted first physical quantity (PQ1 E ; PQ1 D / F ) to the first machine learning model (30; 34) together with the input data not including the measured first physical quantity (RI) for obtaining modified output data (OD*) for use in detecting a fault condition of the machine, the modified output data (OD*) comprising at least one adjusted estimated second physical quantity (PQ2 E ), in which at least one adjusted estimated second physical quantity (PQ2 E ) the external influence has been limited.

15. A computer program product for adjusting data used for detecting a fault condition of a machine (10), the computer program product comprising a data carrier (26) having the computer program code (24) according to claim 14.

Images