Method for providing a classifier or regressor for diagnosing a machine

Abstract

The invention relates to a method for providing a classifier or regressor for diagnosing a machine (20) of a second type. This assumes the availability of an already trained classifier or regressor for diagnosing a machine (10) of a first type. The derivation of the classifier or regressor for diagnosing the machine (20) of the second type is then carried out with comparatively little effort.

Description

Prior ArtFor example, from the publication "'New approach to intelligent fault diagnosis of rotating machinery', Yaguo Lei et al., Expert Systems with Applications 35, (2008) 1593-1600", it is known to examine machines for possible malfunctions on the basis of sensors, for example acceleration sensors.From the applicant's post-published DE 10 2022 210 047 A1, among other things, a method for training a classification algorithm for assigning a machine to one of several damage classes is known, wherein the method for training comprises the following steps:receiving training data comprising reference signals from operated reference machines, which are in particular structurally identical to the machine, wherein one of the damage classes is assigned to each individual reference machine by the training data;determining reference values of one feature or of a plurality of features from the reference signals,training the classification algorithm on the basis of the reference values determined for the feature of the reference machines and on the basis of the damage classes assigned to the reference machines.Disclosure of the InventionThe present invention is based on the observation by the inventor that the training of the classification algorithm is very complicated, since a multiplicity of damage images on real reference machines actually have to be prepared in order to obtain corresponding training data, and since the application of the trained classification algorithm is subsequently restricted to machines which are structurally identical to the reference machines, while the diagnosis of other machines which are not structurally identical to the reference machines requires separate training data and separate training.The invention now follows the approach of starting from an already existing trained classifier or regressiver for diagnosing a machine of a first type of construction, wherein the classifier or regressiver assigns a damage class of a plurality of damage classes, which identify a state of the machine of the first type of construction, or assigns a value of a variable, which identifies a state of the machine of the first type of construction, starting from signals measured on the machine of the first type of construction and transformed into the frequency space with at least one first acoustic sensor.The invention follows the objective of getting with little further effort to a classifier or regressiver for diagnosing a machine of a second type, wherein the first type can be different from the second type and wherein the classifier or regressiver assigns a damage class of a plurality of damage classes, which identify a state of the machine of the second type, or assigns a value to a variable, which identifies a state of the machine of the second type, starting from signals measured on the machine of the second type with at least one second sound sensor and transformed into the frequency space.For the solution, the following method steps are provided first:providing first excitation data representing a measurement of sound in the frequency space with the at least one first sound sensor, wherein the sound is generated by an operated machine of the first type, wherein a first specially prepared rotating machine element is integrated into the machine of the first type,providing second excitation data representing a measurement of sound with at least one second sound sensor in the frequency space, wherein the sound is generated by an operated machine of the second type, wherein a second specially prepared rotating machine element is integrated into the machine of the second type,calculating a normalization function from the first excitation data and the second excitation data.The classifier or regressive for diagnosing the machine of the second type then works in the following steps:providing signals measured by the at least one second sound sensor on the machine of the second type and transformed into the frequency space;calculating a normalized signal from the normalization function and the signals measured with the at least one second acoustic sensor and transformed into the frequency space;associating a damage class of a plurality of damage classes that identify a state of the second-type machine or a value of a variable that identifies a state of the second-type machine with the second-type machine, starting from the normalized signal with the trained classifier or regressive for diagnosing the first-type machine.In this case, the classifier or regressive device can be used for diagnosing the machine of the first type of construction directly, that is to say without further learning.In other words, the present invention is based on the approach of continuing to use a trained model already existing for a first machine type and the knowledge contained therein for a second machine type, and to perform only an adaptation on the basis of two comparatively simple measurements and simple mathematical operations. The effort required to provide the second model is thus significantly reduced compared to the effort which was generally required to provide the first model.The property "structurally identical" is understood here as throughout the following: Two objects are structurally identical if they differ from one another only insofar as it can be expected from the perspective of the person skilled in the art in the case of unintentional production-related copy scattering. If the two objects differ by more than this, they have in particular different geometries, dimensions, materials and / or joints, then their designs are different, and they are therefore not of the same construction.Signals or functions in the frequency space are understood in the present case to mean data which respectively associate a value with a set of frequency values, which value can be, for example, an amplitude and or a power density. Operations such as division or multiplication between signals or functions are then performed by performing the corresponding operations between the values of the signals or functions associated with the same frequency values.The first / second excitation data, which represent a measurement of sound with the at least one first / second sound sensor in the frequency space, can actually be attributed to physical measurements with sound sensors. The first / second excitation data, which represent a measurement of sound with the at least one first / second sound sensor in the frequency space, can, however, also be just the result of a simulation calculation.It can advantageously be provided that the normalization function is calculated as a quotient from the second excitation data and the first excitation data, according to whichN(f) is the normalization function,a2(f) the second excitation data anda1(f) is the first excitation data.It can advantageously be provided that the normalized signal is calculated as a product of the normalization function using the at least one second acoustic sensor and transformed into the frequency space, according to whichn2(f) the normalized signal,N2(f) the normalization function andg2(f) is the signal measured by the at least one second sound sensor and transformed into the frequency space.Alternatively and encompassed by the invention according to claim 1, is a procedure in which the classifier or regressiver has already been trained with standardized training data for the diagnosis of a machine of a first type, for example according to whereinn1(f) are normalized training data,g1(f) are training data anda1(f) is the first excitation data.In this special case, the normalized signal can then be determined directly according ton2(f) is the normalized signal,g2(f) is the signal measured by the at least one second sound sensor and transformed into the frequency space, anda2(f) is the second excitation data.It can advantageously be provided that the at least one first and the at least one second acoustic sensor are identical to one another or are identical in construction to one another.Advantageously, the at least one first and the at least one second sound sensor can be structure-borne sound sensors."At least one" sound sensor comprises the possibility that expressly only a single sound sensor is provided, but it also comprises the possibility that a plurality of sound sensors are provided. Properties which are characterized as properties of the "at least one" sound sensor are then properties which necessarily have to be assigned to each of the plurality of sound sensors or else can have only individual ones of the plurality of sound sensors.It can be provided that the first specially prepared rotating machine element and the second specially prepared rotating machine element are identical to each other. For example, it can be a machine element stored for the purpose of generating excitation data, which is integrated into a machine for generating excitation data and removed from the machine after the generation of excitation data and replaced by another machine element, which is not necessarily specially prepared. The specially prepared rotating machine element is then subsequently available for integration into further machines.The special preparation can aim to excite vibrations or sound significantly on the machines with the specially prepared rotating machine elements, for example to excite structure-borne sound in the machines. The special preparation can correspond, for example, to a wear pattern of the rotating machine elements such as surface removal or the like.In this respect, the first specially prepared rotating machine element and the second specially prepared rotating machine element can also be machine elements of identical construction which are prepared in the same manner.It may be provided that the first specially prepared rotating machine element is integrated into the first machine in a manner corresponding to its nature in the manner in which the second specially prepared rotating machine element is integrated into the second machine, and wherein the manner in which the first sound sensor measures sound of the first machine corresponds to its nature in the manner in which the second sound sensor measures sound of the second machine. This can be done on the basis of structural features which are common to the machine of the first type and to the machine of the second type. If, for example, the machine of the first type is an electric motor with a rotor shaft which is mounted in a bearing, and if, for example, the machine of the second type is another electric motor, likewise with a rotor shaft which is mounted in a bearing, then it can be provided that the respective specially prepared rotating machine element is integrated instead of the bearing of the first or second machine for the purpose of providing the excitation data.It can be provided that, for obtaining the first excitation data and for obtaining the second excitation data, the machine of the first type and the machine of the second type are operated with comparable parameters, for example with the same rotational speed of the rotating machine element. This is in particular a rotational speed of the rotating machine element, which is relevant for the state monitoring of the machine of the second type. On the other hand, however, it is also possible for this rotational speed of the rotating machine element to deviate from typical operating conditions during the state monitoring of the machine of the second type of construction or even to deviate greatly.Under these conditions, the method according to the invention can also be understood to mean that the diagnosis of a machine of a second type relates in particular to the diagnosis of a rotating machine element of the machine of the second type, wherein the diagnosed rotating machine element of the machine of the second type is integrated into the machine of the second type in the same way as is the second specially prepared rotating machine element for providing the second excitation data.It can be provided that the manner in which the first sound sensor measures sound of the first type machine corresponds to its nature according to the manner in which the second sound sensor measures sound of the first type machine.In the above example, for example, for the first acoustic sensor and the second acoustic sensor, structurally sound sensors of identical construction can be used, which are mounted on the outside on a motor housing at the axial height of the above-mentioned bearings.It can be predefined, for example, as a boundary condition of the present methods that machines of the first type match machines of the second type, even if they are not of identical construction to one another, but with regard to their rotating machine elements, e.g. roller bearings, and / or with regard to their loads and / or with regard to their sensors.The invention also relates to a method for diagnosing a machine of a second type having a classifier or a regressive device which has been provided by a method as described here.Although the rotating machine element may be a rolling bearing, the invention is not limited thereto. The rotating machine element can also be a gear toothing, for example.The above-mentioned trained classifier or regressive for the diagnosis of a machine of a first type is assumed to be provided for the method according to the invention.With regard to the properties of the trained classifier or regressive for the diagnosis of a machine of a first design, reference can be made to the post-published DE 10 2022 210 047 A1 of the applicant.In this respect, this trained classifier for diagnosing a machine of a first type or the method for diagnosis provides the assignment of a machine of the first type to one of a plurality of damage classes. The plurality of damage classes can comprise, for example, a class which corresponds to a machine of the first type which operates without faults on the basis of objective criteria. The plurality of damage classes can comprise, for example, one or more classes, corresponding to machines of the first type which do not operate without errors on the basis of objective criteria, or have a fault pattern specific to the respective class.Such a regression algorithm for diagnosing a machine of a first design can, for example, quantitatively determine a remaining service life of the machine, within which the machine is likely still to operate without errors.The machine of the first type of construction has, for example, a rotor mounted in a rolling bearing and at least one stator coil.The above-mentioned defects or damage can be based, for example, on mechanical damage to the rolling bearing, for example failure of lubrication due to lubricant aging, material fatigue with cut-outs on running surfaces, fractures in rolling body cages and the like.The assignment of the machine of the first type to a damage class is carried out, for example, for the diagnosis of a machine of a first type, wherein the assignment can be carried out by means of a machine learning algorithm; it can be, for example, a neural network or, for example, a support vector machine (Support Vector Machine), for example, a quadratic support vector machine.The assignment of the machine of the first type to a damage class is moreover carried out here, for example, on the basis of one feature or a plurality of features or transformed features. More precisely, the assignment takes place, for example, on the basis of a specific value of the feature or on the basis of specific values of the features or of the transformed features, wherein the determination of the value / values takes place on the basis of a signal previously received by the sensor.Such features are defined in the frequency range in the present case. These may be, for example, the following features proposed in the publication mentioned at the beginning:Here, s(k) is the discrete frequency spectrum associated with a time-sampled signal and f k is the frequency associated with s(k).It is also possible to assign the machine of the first type to a damage class directly on the basis of the raw data in the frequency range s(k). This is preferably considered in methods of deep learning, for example in so-called convolutional networks.A transformation from the time domain into the frequency domain (in the present case also: frequency space) can be carried out in an efficient manner, for example, by means of Fast Fourier Transformation (FFT).As already explained, the assignment of the machine of the first type to a damage class can be effected not only on the basis of one feature or more features, but also on the basis of transformed features which are derived from the features by a transformation, in particular a linear transformation, preferably a principal axis transformation or a principal component analysis (PCA).For example, it is provided that the classification algorithm is trained for diagnosing a machine of a first type on the basis of reference machines, for example on the basis of an ensemble of reference machines, wherein each individual reference machine of the ensemble is assigned to one of the damage classes. It can be provided that at least one of the reference machines is also assigned to each damage class.The ensemble of reference machines can comprise, for example, at least 10 or at least 100 individual reference machines. The reference machines can be different from one another over their full circumference. However, it can also be provided that reference machines are only partially different from one another, for example, by exchanging only one component, for example, a rolling bearing, merge into one another, wherein other components, for example, a rotor and / or a stator, can remain identical.The reference machines are in particular identical in construction to the machine of the first type of construction, i.e. they differ from it and from one another only insofar as it corresponds to the possible individual damage and to a production- and aging-dependent copy scatter.The training of the classification algorithm for diagnosing the machine of the first type takes place in particular by operating the reference machines and receiving the reference signals associated with each reference machine.A reference value of the feature can then be determined from each reference signal for the feature or for each feature.Finally, the training of the classification algorithm for diagnosing the machine of the first type can be carried out on the basis of the reference values determined for the feature of the reference machines and on the basis of the damage classes assigned to the reference machines.In the variant already mentioned above, in which the assignment of the machine of the first type of construction is carried out on the basis of transformed features, the training of the classification algorithm is also carried out on the basis of the transformed features and the associated transformed reference values. As already explained, the transformed features can be derived from the features by a linear transformation, in particular by means of a principal component analysis, which is carried out on the basis of the features and the associated reference values and the damage classes assigned to the reference machines.The transformation, or the linear transformation, in particular the principal component analysis, can be used to reduce the number of features used for the state determination and thus to minimize a computing effort required for the method. For example, the number of transformed features may be not greater than 10 or not greater than 5. The method can then be carried out more quickly, more robust and transparently.An operated reference machine or an operated machine of the first type is distinguished, for example, in that it is an electric machine, wherein its rotor is mounted in its rolling bearing and rotates relative to its stator coil.The method according to the invention can be implemented on a computer. Accordingly, the invention also comprises a computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method according to the invention, and a computer-readable data carrier on which this computer program is stored. Finally, the method according to the invention also comprises a computer which comprises such a computer-readable data carrier and an evaluation and control unit which comprises such a computer-readable data carrier and furthermore comprises means for carrying out the method steps according to the invention.The methods according to the invention can be carried out in connection with machines of the first design, otherwise of all types, in particular with electric machines. The latter can in turn be arranged, for example, in machine tools, in commercial vehicles, for example, in off-road vehicles, on wheel bearings of trailers and E-axles for any vehicles or else washing machines, battery-powered screwdrivers or the like.The drawing shows:FIG. 1 schematically shows a first machine with a bearing and a sound sensor;FIG. 2 schematically shows a second machine with a bearing and a sound sensor;FIG. 3 shows a flow diagram of an exemplary embodiment of the method according to the invention for providing a classifier or regressive for diagnosing a machine;FIG. 4 shows a flow diagram of a first classifier or regressive for the diagnosis of a machine.FIG. 1 schematically shows an electric machine 10 of a first design having a rolling bearing 12 and a structure-borne sound sensor 13.For this electric machine 10, the existence of a trained classifier or regressiver for its diagnosis is presumed, wherein the classifier or regressiver, starting from signals measured with at least the structure-borne sound sensor 13 and transformed into the frequency space, assigns a damage class of a plurality of damage classes, which identify a state of the machine 10, or assigns a value to a variable, which identifies a state of the machine 10 of the first type of construction. This classifier or regressive is provided in method step S 1 (see FIG. 3 ).FIG. 2 schematically shows an electric machine 20 of a second design having a rolling bearing 22 and a structure-borne sound sensor 23.While the type of the rolling bearings 12, 22 of the first type of electric machine 10 and the second type of electric machine 20 are identical in the example, the structure-borne sound transfer paths 15, 25 of the first type of electric machine 10 and the second type of electric machine 20 are different from one another in the example.In order to provide a classifier or regressive for diagnosis of the machine 20 of the second type of construction (see FIG. 2 ) with little outlay, the following further steps are provided, see FIG. 3.Method step V 2: In the machine of the first type 10, see FIG. 2, a specially prepared rolling bearing is integrated as rolling bearing 12. The specially prepared rolling bearing has, for example, bearing damage, so that it vibrates greatly during operation. During operation of the machine 10 of the first design, structure-borne sound thus reaches the structure-borne sound sensor 13 along the structure-borne sound transfer path 15.Method step V 3: The specially prepared rolling bearing is now integrated as rolling bearing 22 into the machine of the second type 20, see FIG. 3. In the example, the specially prepared rolling bearing previously used in method step V 2 in the machine of the first type 10 (and removed there in the meantime) is used identically. During operation of the machine 20 of the second design, structure-borne sound now likewise reaches the structure-borne sound sensor 23 along the structure-borne sound transfer path 25, The signals sensed by the structure-borne sound sensor 23 are digitized and transformed into the frequency space using an FFT and subsequently provided as second excitation data a 2( f) for further processing.Method step V 4: By dividing the second excitation data a 2( f) element-by-element (i.e. data assigned to the same frequencies in each case) by the first excitation data a(f), a normalization function N 2( f) is calculated as the resulting quotient:Subsequently, in method step V 5, the classifier or regressive for diagnosis of the machine 20 of the second type is set up to operate in the following classification or regression steps KR 1 to KR 3 (see also FIG. 4 ):Classification or regression step KR 1: Providing a signal g 2( f) measured by the structure-borne sound sensor 23 on the machine of second design 20 and transformed into the frequency space. In this case, instead of the rolling bearing 22, any rolling bearing 22 to be diagnosed is provided.Classification or regression step KR2: calculating a normalized signal n2(f) as a product of the normalization function N2(f) and the signal g2(f) measured with the at least one second sound sensor and transformed into the frequency space:Classification or regression step KR 3: Assigning a damage class of a plurality of damage classes which identify a state of the second-type machine 20 or a value of a variable which identifies a state of the second-type machine 20 to the second-type machine 20, on the basis of the normalized signal n 2( f) using the trained classifier or regressive for diagnosis of the first-type machine 10.In this case, the trained classifier or regressive device can be used directly for diagnosing the machine 10 of the first type in a first variant.References included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedDE 10 2022 210 047 A1 [0002, 0032]Cited Non-Patent Literature'New approach to intelligent fault diagnosis of rotating machinery', Yaguo Lei et al., Expert Systems with Applications 35, (2008) 1593-1600

[0001]

Claims

Method for providing a classifier or regressiver for diagnosing a machine (20) of a second type, by the following steps: - providing a trained classifier or regressiver for diagnosing a machine (10) of a first type, wherein the classifier or regressiver assigns a damage class of a plurality of damage classes, which identify a state of the machine (10) of the first type, or assigns a value of a variable, which identifies a state of the machine (10) of the first type, to the machine (10) of the first type, starting from signals measured on the machine (10) of the first type with at least one first sound sensor (13) and transformed into the frequency space, - providing first excitation data (a1(f)), which represent a measurement of sound with the at least one first sound sensor (13) in the frequency space, wherein the sound is generated by an operated machine (10) of the first type, wherein a first specially prepared rotating machine element is integrated into the machine (10) of the first type, - providing second excitation data (a2(f)) which represent a measurement of sound with at least one second sound sensor (23) in the frequency space, wherein the sound is generated by an operated machine (20) of the second type, wherein a second specially prepared rotating machine element is integrated into the machine (20) of the second type, - calculating a normalization function (N2(f)) from the first excitation data (a1(f)) and the second excitation data (a2(f)), wherein the classifier or regressiver is configured for diagnosing the machine of the second type, To operate in the following steps: - providing signals (g2(f)) measured with the at least one second acoustic sensor (23) on the machine (20) of second type and transformed into the frequency space); - calculating a normalized signal (n2(f)) from the normalization function (N2(f)) and the signals measured with the at least one second acoustic sensor (23) and transformed into the frequency space; - assigning a damage class of a plurality of damage classes which characterize a state of the machine (20) of second type or a value of a variable which characterizes a state of the machine (20) of second type to the machine (20) of second type on the basis of the normalized signal (g2(f)) with the trained classifier or regressiver for diagnosing the machine of first type.Method according to Claim 1, wherein the normalization function (N(f)) is calculated as a quotient from the first excitation data (a1(f)) and the second excitation data (a2(f)).Method according to either of Claims 1 and 2, wherein the normalized signal (n2(f)) is calculated as a product of the normalization function (N2(f)) with the signal (g2(f)) measured with the at least one second acoustic sensor (23) and transformed into the frequency space.Method according to Claim 1, wherein the classifier or regressiver has already been trained for the diagnosis of a machine of a first type with normalized training data (n1(f)) which emerge from training data (g1(f)) by division with the first excitation data (a1(f)), and wherein the normalized signal (n2(f)) emerges from the signal (g2(f)) which is measured with the at least one second acoustic sensor (23) and transformed into the frequency space by division with the second excitation data (a2(f)).Method according to one of Claims 1 to 4, wherein the at least one first and the at least one second acoustic sensor (13, 23) are identical to one another or are identical in construction to one another.Method according to one of Claims 1 to 5, wherein the at least one first and the at least one second sound sensor (13, 23) are structure-borne sound sensors.Method according to one of Claims 1 to 6, wherein the first specially prepared rotating machine element and the second specially prepared rotating machine element are identical to one another or are identical in construction to one another.The method of any one of claims 1 to 7, wherein the first specially prepared rotating machine element is integrated into the first type machine (10) in a manner corresponding to its nature in the manner in which the second specially prepared rotating machine element is integrated into the second type machine (20), and wherein the manner in which the first acoustic sensor measures sound of the first type machine (10) corresponds to its nature in the manner in which the second acoustic sensor measures sound of the second type machine (20).The method according to any one of claims 1 to 8, wherein the first specially prepared rotating machine element and the second specially prepared rotating machine element are rolling bearings (12, 22).Method according to one of Claims 1 to 9, wherein the diagnosis of a machine (20) of a second type relates to the diagnosis of a rotating machine element of the machine (20) of the second type, wherein the diagnosed rotating machine element of the machine (20) of the second type is integrated into a machine (20) of the second type in the same way as the second specially prepared rotating machine element is for providing the second excitation data (a2(f)).The method of any one of claims 1 to 10, wherein the first type is different from the second type.Method for diagnosing a machine of a second type with a classifier or a regressive device which has been provided with a method according to one of the preceding claims.

Images