Model of a soft sensor for measuring mechanical damage, and a soft sensor

Abstract

The invention relates to a method for providing a model (101) of a soft sensor for measuring mechanical damage to an electric rotating machine (103), wherein the model (101) is contained in the soft sensor, - acceleration sensor data (106) measured at the electric rotating machine (103) and magnetic field sensor data (107) corresponding to the acceleration sensor data (106) are received via a first interface (104), - an operating point (111) of the electric rotating machine (103) is determined *from the magnetic field sensor data (107) by a computer device (105), acceleration sensor data (106) corresponding to the operating point (103) is determined and assigned to the operating point (111), and a severity of damage to a mechanical component of the electric rotating machine (103) - a damage severity (114) - is determined from the acceleration sensor data (106) corresponding to the operating point (111), a model (101) is formed *based on the magnetic field sensor data (107), the acceleration sensor data (106), the operating point (111) and the damage severity (114), the model receiving magnetic field sensor data and acceleration sensor data (106, 107) as input variables and outputting operating points (111) and the damage severity (114) as output variables, - the model (101) is provided via a second interface (120).

Description

[0001] Description

[0002] Model of a soft sensor for measuring mechanical damage and a soft sensor

[0003] The present disclosure relates to a method for providing a model of a soft sensor, the model being included in the soft sensor, for measuring mechanical damage of an electric rotary machine, a soft sensor having such a model, and a method for measuring the mechanical damage of the electric rotary machine by means of such a soft sensor.

[0004] US 2022 / 011763 A1 discloses: A method for monitoring at least one machine, comprising: causing at least one first sensor to detect at least one first non-stationary signal from at least one machine operating in a non-stationary manner during at least one operating period, wherein the at least first sensor provides at least one first non-stationary output; causing at least one second sensor to detect at least one second non-stationary signal from the at least one machine during the operating period, wherein the at least second sensor provides at least one second non-stationary output; fusing the at least first non-stationary output with the at least second non-stationary output to produce a fused output;Extracting at least one feature from at least one of the first and second non-stationary signals based on the fused output, analyzing the at least one feature to determine a health status of the at least one machine, and performing at least one of the following steps: repair operation, maintenance operation, and changing operating parameters of the at least one machine based on the health status determined by the analysis. WO 2022 / 148751 A1 relates to a sensor device for monitoring a rotating machine with a motor. The sensor device is suitable for attachment to a surface of the rotating machinery. The sensor device comprises an outer housing with an outer surface. The sensor device is for monitoring the condition of the rotating machinery using a magnetic field sensor.which is configured to measure a magnetic field during the rotation of the rotor, an accelerometer which is configured to measure mechanical vibrations during the rotation of the rotor, and / or a measuring circuit comprising a first sensor element and a second sensor element which are configured to measure a field during the rotation of the rotor. The sensor device is configured with a data processor arrangement and one or more signal units in communication with the data processor arrangement. The data processor arrangement is configured to process inputs based on measurements taken and according to a predefined set of rules. The sensor device is configured to provide a summarized output based on the processing,which indicates a machine state selected from a plurality of machine states. The disclosure also relates to corresponding methods for monitoring rotating machines. Computer programs for carrying out the method steps and a computer-readable medium for storing the computer programs are also disclosed.

[0005] When monitoring the condition of electrical drives, such as electrical rotating machines, and their industrial applications, one of the most important measured variables is vibration. If measured using sufficiently good measuring technology and in sufficient proximity to the mechanical components, it is a good indicator of possible faults and damage. One example of a fault that can be detected by vibration measurement in electrical rotating machines is bearing damage and other mechanical faults, such as misalignment of the axles or mechanical imbalances in the machine or application, as well as loose foundations. A fundamental problem here is that vibration measurements can depend on many external factors, particularly if measurements are not taken directly on the mechanical components (e.g. bearing housings).In particular, the load condition of the machine or its application has a significant influence on the vibration amplitude as well as the characteristics of the damaging frequencies in the spectrum, as well as the speed.

[0006] The central problem in assessing mechanical damage, for example the damaged condition of a bearing or the severity of the damage, is that this cannot be assessed solely on the basis of the occurrence of specific damage frequencies, as it has been shown that these can occur even in a newly installed, fault-free bearing. Likewise, measurements taken with an external sensor, i.e. measurements not taken directly on the bearing housing, cannot provide a generally valid assessment of the severity of damage based on frequency characteristics. In order to make a general statement regarding bearing damage or other mechanical damage occurring at a specific time, the individual measurements must be related to one another.

[0007] Common measurement systems for condition monitoring are either used on a random basis to obtain a short-term overview of the health of the bearing and the mechanical components, or they evaluate individual, regularly transmitted, high-resolution vibration or acoustic data with regard to characteristics of bearing damage or the damage conditions mentioned above. The evaluation is usually based on individual raw data recordings, each of which is examined more or less independently of one another. This usually involves filtering the frequency spectrum according to the frequency components of the mechanical rotation frequency or the known bearing damage frequencies, and an assessment of the severity of possible damage is made based on these frequency components of the overall spectrum. A major disadvantage of most common systems is that precise detection and assessment of mechanical damage, e.g.of bearing damage, specific damage frequencies of the mechanical components in question (e.g. in the case of a bearing: inner ring frequency, outer ring frequency, rolling element frequency, cage frequency, etc.) and therefore also their geometry and their specific type, manufacturer must be known - so in the case of a bearing: bearing geometry, bearing type, and manufacturer. In practice, this specific information is often no longer available after the mechanical component has been replaced and is also not easily obtainable for an installed component without it being removed again.

[0008] Due to the above, there is a need for a soft sensor that can determine damage severity of mechanical components with little prior knowledge about these mechanical components.

[0009] For this purpose, a method of the above-mentioned type is proposed, in which acceleration sensor data measured on the electric rotary machine and magnetic field sensor data corresponding to the acceleration sensor data are received via a first interface, by a computing device,

[0010] * an operating point of the electric rotary machine is determined from the magnetic field sensor data, acceleration sensor data corresponding to the operating point are determined and assigned to the operating point, and a severity of damage to a mechanical component of the electric rotary machine - a damage severity - is determined from the acceleration sensor data corresponding to the operating point,

[0011] * based on the magnetic field sensor data, the acceleration sensor data, the operating point and the damage severity, a model is formed which receives magnetic field sensor and acceleration sensor data as input variables and outputs the damage severity as output variable, the model is provided via a second interface.

[0012] In other words, the present disclosure relates to a method for creating or training a soft sensor for or on a (concrete) electric rotary machine.

[0013] The damage severity is determined as a function of the time during which the acceleration sensor data was measured. The model preferably provides the damage severity as a function of time—that is, a temporal trend of the damage severity—as its output variable.

[0014] It is understood that different operating points can be associated with different levels of mechanical damage.

[0015] In one embodiment, it can be provided that the acceleration sensor data and the magnetic field sensor data are measured with a hardware sensor contained in the soft sensor, which is arranged, in particular attached, to the outside of the machine housing, for example. In particular, it can be provided that the hardware sensor is not arranged on the mechanical component, for example, not on the bearing.

[0016] In one embodiment, it can be provided that the acceleration sensor data is pre-processed. The pre-processing can, for example, comprise a correction of the time stamps assigned to the data and / or filtering according to specific frequencies. Bearing damage frequencies can, for example, be in a range between 3 kHz and 8 kHz. In one embodiment, it can be provided that the acceleration sensor data and the magnetic field sensor data are measured over a large number of preferably disjoint or non-overlapping time intervals of a predetermined length, which is preferably 5 to 10 seconds. Such individual measurements can take place at different times, for example 2-3 times a day, several days in succession, 10 to 30, in particular 15 days. This produces a large number of raw data snippets, each snippet containing data from an individual measurement, for example 10 seconds long.The snippets are preferably high resolution.

[0017] The damage severity can be determined for individual, or preferably for all, accelerometer data snippets. For example, the damage severity calculated for a single accelerometer data snippet corresponds to a damage severity data point. Several such damage severity data points can form the damage severity as a function of time—i.e., its temporal development or trend.

[0018] In other words, the accelerometer data is collected, preferably over a longer period of time.

[0019] Thus, in one embodiment, it can be provided that the acceleration sensor data and the magnetic field sensor data comprise a plurality of data snippets, wherein each acceleration sensor data snippet is related to (and thus corresponds to) exactly one magnetic field sensor data snippet.

[0020] The measurement intervals for the acceleration sensor data and the magnetic field sensor data do not have to be identical. This means that the acceleration and magnetic field sensor data snippets can be configured differently—have different sizes, contain different amounts of data, etc. For example, the acceleration sensor data can contain vibration values ​​measured at 6.6 kHz in the X, Y, and Z directions for 6 seconds. A magnetic field measurement, for example, can be taken at 500 Hz for two minutes.

[0021] In one embodiment, it can be provided that one or more damage indicators are determined from the acceleration sensor data by a computing device and the severity of the damage is determined from the one or more damage indicators.

[0022] Preferably, the model is configured to output the damage indicator(s) as output variable(s).

[0023] It may be useful if the one or more damage indicators characterize a proportion and / or a characteristic of cyclostationary frequencies, i.e. the amplitude of the signals corresponding to the cyclostationary frequency, in the, for example, preprocessed acceleration sensor data, preferably in the respective acceleration sensor data snippet.

[0024] In one embodiment, it can be provided that the one or more damage indicators are calculated with the aid of an extended envelope spectrum (see Jerome Antoni, Ge Xin, N. Hamzaoui: Fast computation of the spectral correlation). By using the extended envelope spectrum, it is possible to reliably detect cyclostationary frequencies and thus possible damage frequencies even with a low signal-to-noise ratio. This is particularly important for external hardware sensors, since a higher noise level is generally observed here and a larger number of external factors have an additive effect on the signal during the transmission path of the damage signal from the source to the sensor, for example, and thus cause interference.It was recognized that the occurrence of cyclostationary frequencies is an indicator for mechanical damage frequencies, whereby these do not have to be known in detail, but only the approximate frequency range of their occurrence in the specific measuring system (amplification by resonance frequency of the measuring setup, thus stronger occurrence in the specific frequency range, also known frequency range of the occurrence of specific damage according to general knowledge, e.g. for bearing damage between 3 and 8 kHz).

[0025] In one embodiment, the acceleration sensor data may be filtered, for example, using high- or low-frequency filtering. High-pass filtering may be particularly suitable for measuring bearing damage.

[0026] Filtering for cyclostationary frequencies can also increase the signal-to-noise ratio, since the noise is not cyclostationary and is filtered out during filtering.

[0027] In one embodiment, it can be provided that the acceleration sensor data comprise vibration values, preferably vibration values ​​in three mutually orthogonal directions.

[0028] In one embodiment, it can be provided that an operating point model of the electric rotating machine is created on the basis of the magnetic field sensor data, wherein the operating point model characterizes one or more operating points by speed ranges assigned to the operating points, wherein different speed ranges are assigned to different operating points in the operating point model.

[0029] The speed ranges can be determined from the magnetic field sensor data.

[0030] The operating point model can be created, for example, in the cloud. In one embodiment, the operating point model of the electric rotating machine can be created from the (regularly) measured speed and torque values, which are sent, for example, to the cloud and collected there by a cloud application.

[0031] It can be useful to determine the different speed ranges from the magnetic field sensor data and to define the different operating points on the basis of the different speed ranges.

[0032] With regard to the speed ranges, a distinction can be made between the most frequent operating points. For example, a clustering algorithm can be used to make this distinction and create the operating point model.

[0033] In one embodiment, it may be provided that the acceleration sensor data corresponding to the respective operating point are determined using the operating point model. The acceleration sensor data can then be assigned to the respective operating point.

[0034] The acceleration sensor data, preferably all collected acceleration sensor raw data recordings, can be assigned to one of the detected operating points based on their speed or torque.

[0035] For raw data recordings that are not clearly located within a cluster, an assignment is preferably made on the basis of the distance ( e . g . Euclidean distance) to the nearest centroid of one of the operating point clusters .

[0036] After assigning the acceleration sensor data to the respective operating points, whereby preferably each acceleration sensor data recording is assigned to one of the operating points determined, for example as described above, the (preferably all) acceleration sensor raw data recordings and the damage severity and preferably damage KPIs calculated therefrom can each be differentiated according to their operating point.

[0037] In one embodiment, the model may be updated based on further measured accelerometer and magnetic field sensor data, with changes in damage severity (caused by this update) being tracked and stored in the model. An update may, for example, occur monthly, during each maintenance, or similarly.

[0038] In particular, the model can include all past corresponding raw data snippets and the corresponding damage severity, e.g. in the form of damage KPIs, for example since the start of data recording or since the last maintenance interval (e.g. bearing change).

[0039] This update is preferably carried out for all operating points and all damage indicators.

[0040] The model therefore preferably includes damage severity calculated from the raw data snippets for each operating point and thus for the associated ones, e.g., in the form of damage KPIs, as a function of time. In other words, the model enables trend monitoring of damage severity. Damage severity is related, for example, to cyclostationarity, the extent of which can be used to quantify damage.

[0041] The advantage here is that the model does not assess the severity of damage based on a value, but rather on a temporal development.

[0042] In other words, the model can be designed as a trend model or trend detection model. For example, the model can contain an upper limit for the damage severity and, preferably, for the respective damage KPIs.

[0043] In one embodiment, it may be useful if, for the operating point, preferably for each operating point, a trend detection is carried out for the damage severity assigned to the operating point, preferably using a statistical model, for example an ARIMA model (ARIMA stands for autoregressive integrated moving average). This allows the previous trend of the damage severity to be modeled more accurately. If, for example, the frequency intensity becomes stronger, this is an indication that the health status of the electric rotating machine is deteriorating. This can result in messages (e.g., an alarm) being sent to the user earlier, and

[0044] Measures for error prevention (maintenance work such as bearing replacement) can be derived earlier than in the case of the aforementioned limit value being exceeded.

[0045] In one embodiment, it may be provided that a remaining service life of the mechanical component of the electric rotary machine is determined from the changes.

[0046] In other words, it is possible to draw conclusions about the severity of damage as a function of time and / or the approximate remaining lifetime of a mechanical component, e.g. a bearing, from the temporal development / tendency of specific frequency components over a longer period of time.

[0047] The aforementioned need is also met by a method for providing a soft sensor for measuring mechanical damage to an electrical rotary machine, wherein in the method a hardware sensor of the soft sensor is provided, which is intended for measuring the acceleration sensor data and magnetic field sensor data corresponding to the acceleration sensor data on the electrical rotary machine, for example on the machine housing, and a model of the soft sensor is provided as described above.

[0048] Furthermore, the need is met with a soft sensor for measuring mechanical damage of an electrical rotary machine, the soft sensor comprising: a hardware sensor, the hardware sensor being designed and configured to measure acceleration sensor data and magnetic field sensor data corresponding to the acceleration sensor data on an electrical rotary machine, and a model which is provided as described above.

[0049] Such a soft sensor forms an analysis system for detecting and evaluating the severity of an emerging mechanical damage to a mechanical component, in particular a bearing damage.

[0050] In addition, the need is met with a method for measuring mechanical damage to an electrical rotary machine by means of the aforementioned soft sensor, wherein acceleration sensor data and magnetic field sensor data corresponding to the acceleration sensor data, for example in the form of snippets, are measured with the hardware sensor and fed to the model, the model determines a severity of the damage to a mechanical component of the electrical rotary machine from the acceleration and magnetic field sensor data.

[0051] In one embodiment, the model may determine a damage severity as a function of time from the measured data and compare this with the "normal" damage severity provided in the model. The "normal" or non-critical damage severity can, for example, be characterized by a limit value or by a trend that was determined from the measured data when the model was created (see above). Based on the comparison, messages such as alarms can be issued, which can be accompanied by further information for the user (context of the message, etc.).

[0052] For example, if the limit is exceeded multiple times (e.g., on one day), an alarm is triggered. If a limit is exceeded significantly more often on the following day, the situation deteriorates more quickly and becomes more critical.

[0053] If the trend in damage severity has become steeper, the model can check whether this new trend is still within a specified or specifiable tolerance. Depending on the deviation from the previous trend within each operating point, the severity of the deterioration of the mechanical damage pattern, for example of the bearing condition, can be determined. Generally speaking, a steeper increase indicates faster deterioration than the previous trend and therefore greater criticality (maintenance work such as bearing replacement is urgently required), and a flatter increase indicates slower deterioration and therefore less criticality. The result of this evaluation of the damage severity in relation to the previous trend can be communicated to the user, for example via a status change in the form of a traffic light in a dashboard (e.g.Traffic light changes from green to yellow or to red if the trend is very steep) and is preferably used to update the model. Optionally, the user can be simultaneously informed of the urgency of corresponding maintenance measures (e.g., bearing replacement necessary within a few weeks or months).

[0054] In summary, by regularly recording snippets of raw data (e.g., 10 seconds each day) and focusing on the trend or tendency of the calculated damage severity and / or damage indicators (KPIs) instead of the absolute value, a more robust determination of damage severity is possible than with conventional measurement and evaluation methods, such as, in the extreme case, sample measurements, which often cannot guarantee any comparability over time or are only carried out once anyway. Furthermore, evaluating the trend compared to the previous deterioration trend allows for a better forecast of a potential failure and thus of the remaining service life, for example in the case of bearing damage.

[0055] An advantageous embodiment is the following:

[0056] An example of feature extraction is: sensor data is examined within a certain range, for example, using bandpass filtering. This allows peaks and spikes at specific frequencies to be analyzed, as well as ratios to be observed. These can be indicators of damage.

[0057] Fusion advantageously means relating acceleration sensor data with magnetic field sensor data. This advantageously involves combining a time series of acceleration sensor data with a time series of magnetic field sensor data to form a single time series.

[0058] An advantageous embodiment is one in which feature extraction is initially performed separately for the acceleration sensor data. The extracted data is then advantageously merged with the magnetic field sensor data.

[0059] In other words, this can be described as follows: Preferably, feature extraction is performed independently on the individual sensor signals and only then are the magnetic field and the extracted feature vectors fused.

[0060] The invention also relates to a computer-implemented method for providing a model of a soft sensor for measuring mechanical damage in an electric rotating machine. The invention also relates to a computer program product comprising instructions that, when the program is executed by a computing device, cause the computing device to perform the steps of the method.

[0061] Further features, characteristics, and advantages of the present invention will become apparent from the following description with reference to the accompanying figures, which schematically show:

[0062] FIG 1 shows a sequence of a method for providing a model of a soft sensor, and

[0063] FIG 2 shows a sequence of a method for measuring a bearing damage of an electrical rotating machine.

[0064] In the exemplary embodiments and figures, identical or similarly functioning elements may be provided with the same reference numerals. The illustrated elements and their relative sizes are generally not to scale; rather, individual elements may be shown larger in size for clarity and / or clarity.

[0065] FIG 1 schematically shows an exemplary sequence of a method 100 for providing a model 101 of a soft sensor, which in addition to the model 101 comprises a hardware sensor 102.

[0066] The present soft sensor is designed to measure mechanical damage to an electric motor 103 or one or more of its components.

[0067] This damage can, for example, affect the bearing, the misalignment (e.g. on the axle / shaft), the mechanical imbalance, or the loose mounting of the entire motor 103 on the foundation - loose foundation or in English "loosefoot".

[0068] The creation of the model 101 begins by receiving, for example via a first interface 104 of a computing device 105, acceleration sensor data 106 and magnetic field sensor data 107 corresponding to the acceleration sensor data 106.

[0069] The computing device 105 is designed, for example, as a cloud infrastructure and can comprise several computing units that are networked together and work together in a coordinated manner.

[0070] FIG 1 shows that the hardware sensor 102 is attached to the electric motor 103, fixed to the outside of the motor housing, and measures the acceleration sensor data 106 and the magnetic field sensor data 107. The acceleration sensor data 106 can include vibration values, for example, of the electric motor bearing in the X, Y, and Z directions Vib_X, Vib_Y, and Vib_Z.

[0071] The hardware sensor 102 preferably has a communication unit and can be controlled wirelessly, for example, from the cloud infrastructure 105. In particular, the hardware sensor 102 is designed as an IoT device.

[0072] The measurement and upload of the acceleration 106 and the magnetic field sensor data 107 can be done via trigger 108.

[0073] In principle, the trigger 108 can be activated at different times, for example several times (e.g. 2-3 times) per day. A measurement, which can contain a large number of individual measurements, can last a total of several days, e.g. 10, 15, 30 days or longer, and can comprise a large number of triggered measurement sections or individual measurements 109. An individual measurement section 109 can last, for example, a few seconds, preferably 5 to 10 seconds. In particular, the acceleration sensor data 106 can comprise the snippets 106s with vibration values ​​in the X, Y, Z directions, which were measured at 6.6 kHz for 6 seconds.

[0074] For example, a magnetic field measurement can be carried out for two minutes at a frequency of 500 Hz.

[0075] The measurement data (the acceleration sensor data 106 and the magnetic field sensor data 107) can thus contain a plurality of data snippets 106s, 107s.

[0076] Preferably, the computing device 105 relates each acceleration sensor data snippet 106s to exactly one magnetic field sensor data snippet 107s. This establishes a (temporal) correspondence 110 between the acceleration sensor data 106 and the magnetic field sensor data 107.

[0077] The computing device 105 determines operating points 111 of the electric motor 103 from the magnetic field sensor data 107, such that the operating point 111 is known at a time A, B, C, ... that can be determined from the measured data. The computing device 105 can assign the respective operating points 111, for example, to predetermined speed ranges 112. This can be done on the basis of motor characteristics such as the speed and / or torque of the electric motor 103 that are sent regularly, e.g. with each trigger 108, wherein the motor characteristics are determined by the hardware sensor 102. The hardware sensor 102 can be configured to determine speeds and / or torques from the magnetic field sensor data 107.

[0078] Preferably, the operating point 111 is determined for each magnetic field sensor data snippet 107 s.

[0079] The determined operating points 111 can be summarized in an operating point model of the electric motor 103. The operating point model characterizes the operating points 111 by the speed ranges 112 assigned to the operating points 111. In the operating point model, different operating points 111 can be assigned different speed ranges 112.

[0080] For example, the different speed ranges 112 can be determined from the magnetic field sensor data 107. This can be done by the computing device 105 113 or calculated in the hardware sensor 102. Subsequently, the different operating points 111 can be defined based on the different speed ranges 112.

[0081] The operating point model can be created, for example, using a clustering algorithm, which enables a differentiation of the most frequent operating points 111 with regard to the speed ranges 112. An operating point 111 within the framework of the operating point model created using a clustering algorithm is considered to be an operating point cluster that includes several or even a large number of operating points 111 lying within the same speed range 112.

[0082] The computing device 105 determines acceleration sensor data 106 corresponding to the respective operating points and assigns them to the respective operating point 111. This can also be done based on the rotational speed or torque associated with the acceleration sensor data 106.

[0083] In other words, all collected raw data recordings 106, 107 can be assigned to one of the detected operating points 111 based on their speed or torque. The speed or torque can serve as a type of data label.

[0084] The determination of the acceleration sensor data 106 corresponding to the respective operating points (and the assignment) can be carried out using the operating point model. This can be done based on a distance, for example, a Euclidean distance to the nearest centroid of one of the operating point clusters.

[0085] From the acceleration sensor data 106 corresponding to the respective operating point 111, the computing device 105 calculates a severity of the damage to a mechanical component of the engine - i.e. a damage severity 114.

[0086] The damage severity 114 of each operating point 111 can be determined, for example, in the form of one or more damage indicators 115 or damage KPIs (KPI for Key Performance Indicator).

[0087] In a suitable embodiment of the method, the one or more damage indicators 115 can characterize a proportion and / or a characteristic of cyclostationary frequencies in the acceleration sensor data 106, preferably in the respective acceleration sensor data snippet 106s. The use of the cyclostationary frequencies is based on the knowledge that the damage frequencies, particularly in the case of bearing damage, are cyclostationary.

[0088] The damage indicators 115 can, for example, be calculated using an extended envelope spectrum.

[0089] FIG 1 shows that the determination of the damage KPIs 115 from the acceleration sensor data snippets 106s can essentially comprise four steps.

[0090] The correction 116 of the acceleration sensor data 106 or the acceleration sensor data snippets 106s with a more precise sampling rate, or the correction of the timestamps using a corrected sampling rate, and the filtering 117 constitute the steps of data preprocessing. The filtering 117 can be a high- or low-frequency filter and is fundamentally task-specific. For determining bearing damage, for example, high-pass filtering is appropriate because the relevant frequencies lie in the corresponding range of the spectrum.

[0091] It may be expedient to filter the acceleration sensor data 106 according to the cyclostationary frequencies described above. This allows for better filtering out noise and simultaneously improving the signal-to-noise ratio.

[0092] Furthermore, cyclostationary effects depend on the speed, i.e., they are more or less pronounced. For example, the higher the speed, the more pronounced the signal (cyclostationary frequency). In other words, the cyclostationary frequencies exhibit an operating-point-dependent behavior.

[0093] For the acceleration sensor data 106 preprocessed as described above, an extended envelope spectrum can be calculated 118 .

[0094] From the extended envelope spectrum, a calculation 119 of a scalar can be made which indicates the characteristics of cyclostationary frequency components.

[0095] The calculated scalars serve as damage KPIs for the respective operating point 111 .

[0096] In summary, the model 101 is formed 120 based on the magnetic field sensor data 106, the acceleration sensor data 107, the operating point 111, and the damage severity 114. This means, in particular, that the steps described above and carried out by the computing device 105 are present in the model 101 in the form of machine-executable instructions (e.g., computer code), so that the model 101 can be executed by any computing device by processing the instructions present in the model 101. The model 101 receives magnetic field sensor and acceleration sensor data 106, 107 as input variables and outputs the damage severity 114 as output variable.

[0097] It is understood that different operating points 111 can be assigned different mechanical damage severities 114.

[0098] FIG 1 also shows that, if the damage severity 114 is characterized by the damage indicators 115, the model 101 can be configured such that it outputs, for example, the damage severity in the form of the damage indicators 115 assigned to the respective operating point 111 as an output variable.

[0099] The model 101 is provided via a second interface 120 of the computing device 105.

[0100] The Model 101 can be updated using additional measured accelerometer and magnetic field sensor data (not shown here).

[0101] The changes in the damage severity 114 assigned to the operating point 111 caused by this update can be tracked and stored in the model 101. It may be useful to track the changes for all operating points 111 and for all damage KPIs 115.

[0102] The Model 101 update can be performed at regular intervals (e.g., once a month). However, the frequency may depend on an empirical value that provides information about the rate of deterioration of the mechanical component.

[0103] For each operating point 111, a trend can be identified for the relevant damage severity 114. This can be done, for example, using a statistical model, such as an ARIMA model. The remaining service life of the mechanical component of the electric motor can be determined from the changes and, in particular, from the trend of the damage severity 114 assigned to the operating point 111.

[0104] In other words, the updated model describes the previous trend of the calculated damage KPIs 115 within each identified operating point 111 and thus does not depend in principle on the absolute value of the damage KPIs 115, but primarily on their temporal development.

[0105] With Model 101 it is therefore possible to draw conclusions about the severity of a damage and in particular the approximate remaining service life of the mechanical component, for example the bearing and the mechanical components, from the temporal development / tendency of specific frequency components over a longer period of time.

[0106] The method described with reference to FIG. 1 can be viewed as a method for training or creating a soft sensor for a specific electric motor 103. The phase of creating the model 101 represents a phase of familiarizing the electric motor 103 and, in general, an electric rotary machine. Thus, the model describes the normal behavior of the electric motor 103.

[0107] The model 101 can also contain bounds for the damage severity 114. The bounds can be configured as one or more statistical quantities that can be calculated from the damage severity 114, such as a maximum (or another upper limit), standard deviation, etc.

[0108] FIG. 2 schematically shows an exemplary sequence of a method 200 for measuring bearing damage in an electric rotating machine 201. A soft sensor 202 is assigned to the machine 201, wherein a hardware sensor of the soft sensor 202 can be embodied as the hardware sensor 102 of FIG. 1. The hardware sensor 102 is preferably attached to the machine 201 in the same way as to the electric motor 103.

[0109] In addition, the soft sensor 202 has a model 203 .

[0110] The model 203 was created according to the steps described with reference to FIG 1 with the difference that the measurements were carried out on the machine 201, the filtering 117 was designed as a high-pass filter in order to take into account the position of the bearing damage frequencies (e.g. 3 - 8 KHz) and the damage KPIs 115 are thus bearing damage KPIs.

[0111] The model 203 is stored in a computing device 204, which executes the model 203. It is understood that the computing device 105 of FIG. 1 can also be used in the method described with reference to FIG. 2.

[0112] The procedure detects and assesses deviations from the normal condition of the bearing, which is described in Model 203.

[0113] The hardware sensor 102 measures acceleration sensor data and magnetic field sensor data on the electric rotary machine 201. The measurement can be triggered, for example, by a trigger.

[0114] The modalities of the measurement carried out are preferably the same as those described with reference to FIG 1 - duration of a single measurement ( e . g . 10 seconds), its frequency, etc . etc .

[0115] The acceleration sensor data and the magnetic field sensor data are transmitted (preferably in the form of data snippets) to the computing device 204 in a wired or wireless manner and received by the computing device at a first interface 206.

[0116] The model 203 processes the acceleration sensor data snippets 207 and the magnetic field sensor data snippets 208 as described for the acceleration sensor data 106 and the magnetic field sensor data 107 of FIG . 1 .

[0117] Positional damage KPIs 209 are determined from the acceleration sensor data snippets 207 and the respective operating point 210 is determined from the magnetic field sensor data snippets 208.

[0118] This is preferably done, as described above, by determining speed ranges and by clustering the operating points according to the speed ranges.

[0119] It is understood that the (newly determined) situational damage KPIs 209 can be used to update the model 203 .

[0120] The model 203 then carries out a comparison with the data 211 collected during the familiarisation phase of the machine 201 and in this sense describes the normal condition of the bearing.

[0121] For example, model 203 can perform the comparison only based on an upper limit calculation for the respective bearing damage KPIs 212 present in model 203.

[0122] If there are corresponding trends in the model 203 for the bearing damage KPIs 212 stored in the model 203, which were calculated as described above, for example with the help of the ARIMA model, the model 203 can use the same statistical evaluation (e.g. the ARIMA model) to determine deviations from the previous trends.

[0123] Depending on the deviation from the previous trends within an operating point 213 stored in the model 203 corresponding to the determined operating point 210, the model 203 can determine the severity of the deterioration of the bearing condition.

[0124] A steeper increase indicates a faster deterioration than the previous trend and thus a higher criticality (e.g., maintenance work such as bearing replacement is urgently needed). A flatter increase, on the other hand, indicates a slower deterioration and thus a lower criticality.

[0125] The model 203 can communicate the result of this evaluation of the positional damage KPIs 209 calculated from the measurement data 207, 208 with respect to the previous trend to a user. For this purpose, the model 203 can use the computing device 204, which can, for example, include a display device 214. In this case, the model 203 can, for example, check 215 whether certain predetermined threshold values ​​for the specific operating point 210 have been exceeded or whether a trend is apparent.

[0126] For example, the model 203 can cause the computing device 204 to generate a signal that leads to the display of a status change in the form of a traffic light in a dashboard on the display device 214. The traffic light jumps, for example, from green to yellow or to red if the trend increases very steeply. At the same time, the model 203 can cause the computing device 204 to generate another signal that is used to alert the user to the urgency of corresponding maintenance measures (e.g., bearing replacement necessary within a few weeks or months). It is understood that this information (traffic light and indication of urgency) can be contained in the same signal.

[0127] The result of this evaluation of the positional damage KPIs 209 calculated from the measurement data 207, 208 with regard to the previous trend can also be used to update the model 203. The purpose of this description is merely to provide illustrative examples and to indicate further advantages and special features of this invention. Thus, it cannot be interpreted as a limitation of the field of application of the invention or of the patent rights claimed in the claims. In particular, the features disclosed in connection with the methods described here can be usefully used to further develop the sensors described here and vice versa.

Claims

Patent claims 1. A method for providing a model (101) of a soft sensor for measuring mechanical damage of an electrical rotary machine (103), wherein the model (101) is contained in the soft sensor, via a first interface (104) of a computing device (105), acceleration sensor data (106) measured on the electrical rotary machine (103) and magnetic field sensor data (107) corresponding to the acceleration sensor data (106) are received by a computing device (105), * an operating point (111) of the electric rotary machine (103) is determined from the magnetic field sensor data (107), acceleration sensor data (106) corresponding to the operating point (103) are determined and assigned to the operating point (111), and a severity of the damage to a mechanical component of the electric rotary machine (103) - a damage severity (114) - is determined from the acceleration sensor data (106) corresponding to the operating point (111), * based on the magnetic field sensor data (107), the acceleration sensor data (106), the operating point (111) and the damage severity (114), a model (101) is formed which receives magnetic field sensor and acceleration sensor data (106, 107) as input variables and outputs operating points (111) and the damage severity (114) as output variables, the model (101) is provided via a second interface (120) of the computing device (105).

2. The method according to claim 1, wherein the acceleration sensor data (106) and the magnetic field sensor data (107) are measured with a hardware sensor (102) contained in the soft sensor.

3. The method according to claim 1 or 2, wherein the acceleration sensor data (106) and the magnetic field sensor data (107) are measured over a plurality of time intervals of a predetermined length, which is preferably 5 to 10 seconds.

4. The method according to claim 3, wherein the acceleration sensor data (106) and the magnetic field sensor data (107) comprise a plurality of data snippets (106s, 107s, 109), wherein each acceleration sensor data snippet (106s) is related to exactly one magnetic field sensor data snippet (107s).

5. The method according to one of claims 1 to 4, wherein the computing device (105) determines one or more damage indicators (115) from the acceleration sensor data (106) and the damage severity (114) is determined from the one or more damage indicators (115), wherein the model (101) is preferably configured to output the damage indicators (115) as output variables.

6. The method according to claim 5, wherein the one or more damage indicators (115) characterize a proportion and / or a characteristic of cyclostationary frequencies in the acceleration sensor data (106).

7. The method according to one of claims 1 to 6, wherein an operating point model of the electric rotary machine (103) is created on the basis of the magnetic field sensor data (107), wherein the operating point model comprises one or more operating points (111) is characterized by speed ranges assigned to the operating points (111), wherein different speed ranges are assigned to different operating points (111) in the operating point model.

8. The method according to claim 7, wherein the different speed ranges are determined from the magnetic field sensor data (107) and the different operating points (111) are defined on the basis of the different speed ranges.

9. The method according to claim 7 or 8, wherein the acceleration sensor data (106) corresponding to the respective operating point (111) are determined using the operating point model.

10. The method according to any one of claims 1 to 9, wherein the model (101) is updated based on further measured acceleration sensor and magnetic field sensor data (106, 107), wherein changes in the damage severity (114) are tracked and stored in the model (101).

11. The method according to claim 10, wherein for the operating point (111) a trend detection for the damage severity (114) assigned to the operating point (111) is carried out, preferably by means of a statistical model, for example an ARIMA model.

12. The method according to claim 10 or 11, wherein a remaining service life of the mechanical component of the electric rotary machine (103) is determined from the changes.

13. A method for providing a soft sensor for measuring mechanical damage to an electrical rotary machine (103), wherein a hardware sensor (102) of the soft sensor is provided, which is intended to measure the acceleration sensor data (106) and magnetic field sensor data (107) corresponding to the acceleration sensor data (106) on the electrical rotary machine (103), a model (101) of the soft sensor is provided according to a method according to one of claims 1 to 12.

14. Soft sensor for measuring mechanical damage of an electric rotary machine (103), the soft sensor comprising: a hardware sensor (102), wherein the hardware sensor (102) is designed and configured to measure acceleration sensor data (106) and magnetic field sensor data (107) corresponding to the acceleration sensor data (106) on an electrical rotary machine (103), and a model (101) which is provided according to a method according to one of claims 1 to 12.

15. A method for measuring mechanical damage of an electric rotary machine (103) by means of a soft sensor according to claim 14, wherein acceleration sensor data are collected with the hardware sensor (102) (106) and magnetic field sensor data (107) corresponding to the acceleration sensor data (106) are measured and fed to the model (101), the model (101) determines a severity (114) of the damage to a mechanical component of the electric rotary machine (103) from the acceleration and magnetic field sensor data (106, 107).

16. A computer-implemented method for providing a model (101) of a soft sensor for measuring mechanical damage of an electric rotary machine (103), comprising the steps of any one of claims 1 to 12.

17. A computer program product comprising instructions which, when the program is executed by a computing device (105), cause the computing device (105) to carry out the steps of the method according to one of claims 1 to 12.

Images