Method for diagnosing the state of an electric motor

Abstract

The invention relates to a method for diagnosing the state of an electric motor (M) to which an AC feed voltage (U) with a supply frequency (fU) is applied. The method has the following steps: a) detecting a time series (It) of current values of a machine current (I) which flows through the electric motor (M) as a result of the feed voltage (U) applied to the electric motor (M); b) describing the time series (It) using an autoregressive model (formula I) of order p; c) forming a feature vector (Fx) with p vector elements, wherein the p vector elements are formed by the p parameters ai of the autoregressive model; and d) using the formed feature vector (Fx) in an ML algorithm as training data or as test data.

Description

[0001] Description

[0002] Method for diagnosing the condition of an electric motor

[0003] The present invention relates to a method for diagnosing the condition of an electric motor.

[0004] A well-known method for diagnosing the condition of an electric motor is the so-called Motor Current Signature Analysis (MCSA), see, for example, EP3961230A1 (Siemens AG) March 2, 2022 and Kumar, K. Vinoth et al.: A Review of Voltage and Current Signature Diagnosis in Industrial Drives. International Journal of Power Electronics and Drive Systems (IJPEDS) Volume 1, Number 1, September 2011, pages 75 - 82, ISSN 2088-8694. The basis of MCSA is the recording of the electrical phases, usually three (LI, L2, L3), of an electric motor. The recorded

[0005] Measured values ​​are then analyzed: typically, a frequency spectrum is generated from the measured time series using a Fast Fourier Transformation (EFT). MCSA is used to observe changes in the amplitudes of characteristic frequencies in the frequency spectrum in order to detect or quantify faults and operating conditions of the electric motor in the frequency domain.

[0006] The use of EFT involves a conflict between frequency resolution and recording time: On the one hand, high frequency resolution is desired in order to clearly identify motor faults. On the other hand, this requires a relatively long recording time in a constant operating state, i.e., a large amount of measurement data. However, a quasi-stationary operating state over a relatively long period is unrealistic, and on the other hand, handling large amounts of data is complex. Therefore, one object of the present invention is an improved method for monitoring the condition of an electrical machine.

[0007] This object is achieved according to the invention by a method having the features specified in claim 1.

[0008] This is a method for diagnosing the condition of an electric motor. An AC supply voltage with a supply frequency is applied to the electric motor (AC = Alternating Current). In a first step, a time series of current values ​​of a machine current is recorded. This current flows through the electric motor due to the supply voltage applied to the electric motor. The machine current, which flows through the wire coils of the electric motor, generates a magnetic field which, in interaction with a second magnetic field (permanent or electromagnet), causes a rotor of the electric motor to rotate. In a subsequent step, the recorded time series is compared with an autoregression model y t = c + £ t + a i Vt-i of order p : The constant c, the noise term s tand the p parameters a±, which can be called weighting coefficients, are determined, with which a value y t the time series from the p previous values ​​y t -i of the time series, with i = 1, . . . , p . In a subsequent step, a feature vector with p vector elements is created, where the p vector elements are formed by the determined p parameters a± of the autoregression model. In a subsequent step, the created feature vector is used in an ML algorithm as training data or as test data (ML = Machine Learning).

[0009] In the condition diagnosis procedure, the following phases can be run through successively: a Training phase a Recording phase (generation of feature vectors of the "good condition") a2 Learning phase (learning the "good condition" of the electric motor) b Test phase (condition monitoring)

[0010] The invention is based on the finding that it is possible to use autoregression to "condense" the information relevant for anomaly detection from time values ​​of a time series into a few values, so that disturbing information is eliminated, important information is retained, and the amount of data to be processed is considerably reduced.

[0011] The present invention provides a method for diagnosing the condition of an electric motor based on machine learning or AI (artificial intelligence). The input values ​​required for the AI, the so-called features, are obtained from the current samples using an AR model.

[0012] A p-th order autoregression model, notation: AR(p) , can be written as:

[0013] The signal y tconsists of a constant c, a noise term s t and a weighted moving average of the p previous signal values ​​y t -i, where the AR parameters a±, with i = 1, . . . , p, are the weights (AR = autoregression). However, in the context of the present invention, the interest is not in the signal y t itself, but the question of how the AR algorithm processes the signal y t "modeled". The most important input value apart from the signal values ​​y t -i the "lag" or the "order" . The "lag" is the number of periods p that the AR model uses to determine the signal y t used. Has a 1st order autoregressive model = AR ( 1 ) : y t = c + E t + ay t- only a single lag value yt-i, then a 2nd order autoregressive model = AR(2) : already two "lag values" y t -i and y t2 • Autoregressive models of the pth order then have correspondingly p “lag values”.

[0014] AR models are typically used to model discrete time series y(t) , also written as y t to analyze systematic relationships in the past or to make predictions for the future. However, it has been shown that an autoregression model is also very well suited to providing the features for an ML algorithm.

[0015] According to the invention, the parameters a± determined by the AR model form the vector elements of a feature vector that represents the essential information of the input signal, i.e., the time series. In principle, one could also use the time series itself, i.e., the sample values ​​y tof the current, as "features", but this brings with it problems, as explained in the following points a) and b): a) Very large "feature" vectors: With typical sampling frequencies of 2000 to 20,000 sps (sps = samples per second), a recording duration of approximately 1 second results in feature vectors with 2000 to 20,000 values, ie with a size that is difficult to handle. b) Quantity: The current values ​​are quantitatively dependent on the load, ie a higher motor current flows at a higher load. A condition diagnosis method should not, however, classify a load change as an anomaly, but only as an anomaly of the motor.

[0016] The invention circumvents the above-mentioned disadvantage of large data sets, where 2000 to 20,000 current measurement values ​​form a feature vector, by compressing the data using an AR model. Such compression is useful because the information content of a single current measurement value is relatively low and an increase in entropy through autoregression is easily feasible.

[0017] In contrast to a conventional MCSA, the present invention does not use spectral components, i.e., values ​​in the frequency domain, but rather current measurements, i.e., values ​​in the time domain. According to the invention, no physical models are used; instead, the method operates exclusively "generically" using machine learning or AI (= artificial intelligence) methods.

[0018] A further advantage of the method according to the invention is that it solves the problem of "data quality," which is important for ML. In spectral analysis, many features (frequency lines) only represent "noise." On the one hand, these can be static elements, i.e., they do not provide any "useful" information for the system; on the other hand, these "noise" features can lead to completely incorrect models during training because these features have values ​​other than 0. Filtering out these "noise components" in advance is very difficult because, on the one hand, a threshold for the filter would have to be known, and, on the other hand, the frequencies vary depending on the load condition.

[0019] As an alternative to the "standard" autoregression methods, modified AR methods such as ARMA (= AutoRegressive Moving Average), ARMAX (= AutoRegressive Moving Average with exogenous inputs) and ARIMA (= AutoRegressive Integrated Moving Average) can also be used.

[0020] A feature vector generated according to the invention can be used for various AI and ML methods, both for supervised learning methods (e.g. classification and regression algorithms), and for unsupervised learning methods (e.g. unsupervised learning).

[0021] Advantageous embodiments and further developments of the invention are specified in the dependent claims. According to a preferred embodiment of the invention, said electric motor is an asynchronous motor. Due to their cost-effective and robust design, asynchronous motors (= ASM) are very common in industry: they can be found in many applications such as fans, saws and pumps, from relatively small systems in the watt range to systems in the MW range. Asynchronous motors are therefore among the largest consumers of the electrical energy generated worldwide. The widespread use of asynchronous motors offers enormous potential in terms of availability and reliability and the associated savings in time and money.

[0022] According to a preferred embodiment of the invention, the time series of current values ​​is recorded in a phase-locked manner, i.e. the current values ​​are sampled synchronously with the mains frequency, resulting in a constant phase. The current measurement values ​​are therefore always recorded ("sampled") at times at which the current signal has predefined phases, e.g. 64 samples per full wave of the AG current. This phase-locked sampling can significantly improve the condition diagnosis method. The phase-locked sampling can be implemented either using hardware in the form of a PLL (= Phase-Locked Loop). The sampling frequency always adapts to the mains frequency, i.e. the "start phase" is always constant. Alternatively, the current measurement values ​​can initially be sampled at a fixed sampling frequency. In a subsequent step, a "soft PLL" is applied to the current measurement values, i.e. a software-based or...A PLL implemented in the firmware is used. By interpolating the values, a sampling grid is calculated in such a way that a fixed "start phase" is obtained and the sampling frequency is a constant multiple of the mains frequency.

[0023] According to a preferred embodiment of the invention, the method further comprises the following steps: At least one first feature vector is used as training data for the ML algorithm. At least one second feature vector is used as test data for the ML algorithm. A diagnosis of a condition ("condition diagnosis") of the electric motor, which is documented in the form of the second feature vector (test data), is made using the ML algorithm based on a comparison of the at least one second feature vector (test data) with the at least one first feature vector (training data). This embodiment relates to condition monitoring based on an unsupervised learning method. In a first part of a training phase, a so-called recording phase, an electric motor is kept in a "good condition" for a limited period of time, i.e.in error-free operation, during which measured values, e.g. measured values ​​of the current flowing through the motor windings, are recorded and, according to the invention, feature vectors are generated from these, which describe this "good state". In a second part of the training phase, a so-called learning phase, an ML algorithm is trained using several of these feature vectors, i.e. the ML algorithm only sees the "good state" of the electric motor, but does not know the "bad state", i.e. faulty operation of the electric motor. After the training phase has ended, a test phase, also referred to as the detection phase, of the ML algorithm is activated, i.e. from this point onwards the ML algorithm monitors the electric motor for operating states that deviate from the learned "good state" of the electric motor, so-called "anomalies", and reports these to an operator.In detail, the following phases are carried out one after the other: a Training phase a Recording phase (generation of feature vectors of the "good state") a2 Learning phase (teaching the "good state" of the electric motor) b Test phase (condition monitoring) A supervised learning process has the disadvantage compared to an unsupervised learning process that "good" and "bad" data must be trained. This is usually not possible in a customer system. If the training has to be carried out by the manufacturer, a large number of models must be supplied. In "unsupervised" learning, on the other hand, the system is only trained with "good" data; after training, the system recognizes "bad" data as an "anomaly".

[0024] According to a preferred embodiment of the invention, the method comprises the following three steps: it comprises the first step of recording the phase angle at the beginning of the time series for the at least one first feature vector that is used as training data. The phase angle at the beginning of the time series, the so-called “start phase”, is therefore recorded for each time series of current values. The recording of the respective start phase is carried out for all time series in the training phase, and all recorded start phases are documented, e.g. saved as an entry in a list or in a file. The method also comprises the second step of recording the phase angle at the beginning of the time series for at least one second feature vector that is to be used as test data. In the test phase following the training phase, the start phase is therefore also determined from the recorded time series.In addition, the method comprises the third step of only using the at least one second feature vector as test data for the ML algorithm if the phase angle assigned to the second feature vector at the beginning of the time series matches a phase angle assigned to a first feature vector at the beginning of the time series. Therefore, only those time series are evaluated for which there is an entry with the same start phase in the training phase; two phases are considered to be the same if they match within a predetermined tolerance. Therefore, only time series with a “known” start phase are evaluated in the test phase. According to a preferred embodiment of the invention, the data are “internally normalized,” i.e. the phase position and amplitude of the signal are automatically adjusted by the algorithm; this means that there is no effort on the part of the user.The time domain values ​​of the current I are preferably normalized to a maximum value of the absolute value of the current before the AR calculation, ie a measured current value I becomes a normalized current value I. norm = 1 / (max(abs (I) ). This makes the features independent of the signal amplitude. By normalizing the current values, the invention avoids the above-mentioned disadvantage of the quantitative dependence of the current values ​​on the load.

[0025] According to a preferred embodiment of the invention, in addition to the parameters a± of the AR model, information criteria are also calculated, e.g., the Akaike Information Criterion (AIC). The Akaike Information Criterion states that a model should not be unnecessarily complex and should strike a balance between maximum goodness of fit and minimum model complexity. The goodness of fit is a measure of how well the model replicates the real data, e.g., measured by the residual sum of squares. The complexity of the model can be measured, for example, by the number of model parameters. Since the p parameters a± of the AR model are used as vector elements of a feature vector for the ML algorithm, the "lag" determines the size of the feature vector.An information criterion can be used to estimate whether the lag size was chosen appropriately (too large, too small, optimal); based on an information criterion, the lag size can be optimized.

[0026] According to a preferred embodiment of the invention, the feature vector is used for condition monitoring using a self-organizing map (SOM). The idea of ​​the SOM is to generate a grid of X * Y neurons N. Each neuron N contains a weight vector of length L F , i.e., equal to the length of a feature vector. Determining the number of neurons (X * Y) is a compromise between the amount of data / computational effort on the one hand and the desired reliability of an anomaly detector on the other. In previous tests, 20 x 20 = 400 neurons has produced very good results.

[0027] Self-organizing maps, like most artificial neural networks, operate in two modes: training and mapping. In training mode, an input dataset ("input space") is used to generate a map ("map space") as a lower-dimensional representation of the input data. In mapping mode, measurement data is classified using the generated map, for example, as normal or anomalous.

[0028] Learning the SOM, or training, is an iterative process. Initially, the weight vectors contain random values. The learning process then proceeds in such a way that similar feature vectors from the training dataset are assigned to the same neuron. "Similar" is to be understood in the sense of a vectorial distance, i.e., depending on the application, the Euclidean distance, the cosine distance, or the Manhattan distance can be used.

[0029] The goal of training is to represent a p-dimensional input space, i.e., with p variables, as a 2-dimensional map space. A map space consists of components called nodes or neurons, arranged on a hexagonal or rectangular grid with two dimensions. The number of nodes and their position on the grid are determined in advance, based on the goal of analyzing and investigating the input space. Each node in the map space is assigned a weight vector that defines the node's position in the input space. While the nodes in the map space remain anchored to their positions, the weight vectors are approximated to the input data. This minimizes the distance, e.g., the Euclidean distance, between the weight vector and the input data without destroying the topology generated by the map space.

[0030] During the training phase, feature vectors that are "similar" to each other are grouped. This similarity is defined as the Euclidean distance, i.e., the smaller the distance between the feature vectors, the more similar they are. For each "group" of similar feature vectors, one or more representatives are chosen, as is the maximum distance Dmax for each group.

[0031] After completing the training phase, the SOM generated in this way can be used to classify further input space vectors, i.e., observations in the input space, here: time series of current values, by identifying the node whose weight vector is closest to the input space vector, i.e., has the smallest distance metric, e.g., the Euclidean distance.

[0032] A feature vector generated according to the invention with p vector elements, which were obtained by means of an AR method from a time series measured on an electric motor, is assigned to a neuron of the said SOM and to a weight vector assigned to the neuron.

[0033] To use the trained SOM as an anomaly detector, a function is required that assigns a degree to which a feature vector Fv represents an anomaly. For this purpose, the "quantization error" Ev is calculated for each feature vector Fv:

[0034] Ev = I Fv - Gv | for all v, i.e., for all feature vectors. Where | . . . | is the previously defined vectorial distance and Gv is the respective weight vector of the "optimal" neuron, i.e., the neuron whose weight vector G has the minimum distance to the feature vector. The quantization error E describes how well the respective feature vector can be mapped to the distribution of the data with which the SOM was trained. A high error indicates that the feature vector is very different from the training data and therefore describes anomalous motor behavior. In practice, an anomaly in the electric motor is detected if the quantization error Ex exceeds a fixed threshold value E Thr eshoid exceeds: Ex > E Thr eshoid-

[0035] After training, anomaly detection is performed as follows: A feature vector is generated from the input values, as described above. For this vector, the Euclidean distance to the learned "groups" is determined. The minimum value of all distances is determined. If this value is greater than Dmax, an anomaly is present.

[0036] The invention is explained below with the aid of the accompanying drawings. They show schematically and not to scale.

[0037] Fig. 1 shows an embodiment of a drive train with an electric motor;

[0038] Fig. 2 is a flow diagram of an embodiment of the method according to the invention;

[0039] Fig. 3 is a flow diagram of another embodiment of the method according to the invention; and

[0040] Fig. 4 shows an embodiment of a SOM.

[0041] Fig. 1 shows an arrangement with an asynchronous motor 12, which is connected via supply lines 11 to an AC voltage source 10, which has a supply voltage U suppiy with a voltage frequency f suppiy The current I flowing through the supply lines 11 is measured by a current sensor 14, e.g., a shunt, a current transformer, or a Hall sensor. Current measurements are taken at the sampling frequency f A [1 / T] for a measurement period T A [T] is plotted on . The number of samples N A is therefore N A = f A *T A . For example, let f A = 3200 Hz, and T A = 1.28 seconds, thus N A =4096 = 2 12 Samples. At a voltage frequency f suppiy = 50 Hz signal, 3200 / 50 = 64 samples are recorded per AC oscillation, ie 32 samples per AC half-oscillation. The N recorded by the current sensor 14 ACurrent measurement values ​​are sent from the current sensor 14 to a computing unit 16. From these N A The processing unit creates a feature vector for each current measurement value. Lag = 64 is chosen as the minimum lag. This already achieves very good results. This means that the essential information of the time signal can be represented with just 64 values ​​(instead of 4096). Lag = 512 is used as the maximum lag.

[0042] The process steps for determining the feature vectors are performed by the computing unit 16. For this purpose, the computing unit 16 has a processor 18 and a data memory 20. The data memory 20 stores software with an algorithm for generating a feature vector from the current measurement values. This software is executed by the processor 18. Input values ​​such as the measurement duration T Acan be transferred to the computing unit 16 by an input / output unit 22 connected to the computing unit, e.g., a PC. Upon detection of an anomaly, the computing unit 16 sends a corresponding message to the input / output unit 22.

[0043] Fig. 2 shows a flow diagram of an embodiment of the method according to the invention for the condition diagnosis of an electric motor M, to which a supply voltage U with a supply frequency is applied. In a first step 210, a time series I t of current values ​​of a machine current I which flows through the electric motor M due to the supply voltage U applied to the electric motor M.

[0044] In a second step 220, the time series I t with an autoregression model y t = c + E t + a i Vt-i of order p described .

[0045] In a third step 230, a feature vector F xformed with p vector elements, where the p vector elements are formed by the p parameters a± of the autoregression model.

[0046] In a fourth step 240, the formed feature vector Fx is used in an ML algorithm as training data or as test data.

[0047] The recording phase is followed by the learning phase. During this phase, the recorded data and labeled with a slip label L s labeled feature vectors are trained into a SOM.

[0048] An example is described of how the features form the basis for an anomaly detector. The goal here is "unsupervised" learning, meaning the system is trained only with "good" data and should recognize "bad" data as an "anomaly" after training.

[0049] During the training phase, features that are "similar" to each other are grouped. This similarity is defined as follows: Each feature is considered a vector of length L. The similarity is defined as the Euclidian distance, i.e., the smaller the distance, the more similar the features are (https: / / de.wikipedia.org / wiki / Euklidischer_Abstand). For each "group," one or more representatives are chosen, as well as the maximum distance Dmax for each group.

[0050] Fig. 3 shows a flow diagram of an embodiment of the method according to the invention for the condition diagnosis of an electric motor M, to which a supply voltage U with a supply frequency is applied, with the aid of an SOM.

[0051] In a first step 310, current values ​​I m of an electric current flowing through the electric motor M. This gives a time series I t by L F Current values ​​I ma current I , which flows through the asynchronous motor M due to the supply voltage U applied to the asynchronous motor M , while the asynchronous motor M is in a good state . During the recording phase, current values ​​I m recorded, which describe the "good condition" of the motor. These current values ​​I m are then used to train the anomaly detector. The following aspects are very important:

[0052] - The system should cover as many load conditions or slip values ​​as possible in the slip value range [ Smi n , . . . , s N J . For each recording, the respective slip should be recorded as a "label" L s be carried along.

[0053] - If a load condition is recorded, T A the load condition (= slip) should be approximately constant: L s= constant . It makes no sense to record the current values ​​during significant load changes (= slip changes), because this would result in a superposition of different states and thus also a superposition of different frequency spectra.

[0054] - The recording phase is finished when a sufficiently large variation of slip values ​​has been recorded or when the maximum time allotted for recording has been reached.

[0055] In a second step 320, the current values ​​I t A feature vector Fv with p vector elements is formed from the recorded time series, where the p vector elements are formed by the p parameters a± of the autoregression model. A slip value of the slip s of the asynchronous motor is assigned to this feature vector Fv.

[0056] In a third step 330, steps 310 and 320 are repeated for different slack values ​​to obtain a group of feature vectors Fv as a training data set.

[0057] In a fourth step 340, a matrix of X times Y neurons is generated, with each of the neurons being assigned a weight vector G of length L F is assigned .

[0058] In a fifth step 350, those feature vectors Fv from the training data set are assigned to a neuron whose vectorial distance to the neuron does not exceed a predetermined distance limit.

[0059] In a sixth step 360, the weight vectors G of the neurons are adjusted such that the distance of the weight vector G to the feature vectors Fv assigned to a neuron is minimal.

[0060] After training, anomaly detection is performed as follows: A feature vector is generated from the input values ​​as described above. For this vector, the Euclidean distance to the learned "groups" is determined. The minimum value of all distances is determined. If this value is greater than Dmax, an anomaly is present.

[0061] Fig. 4 shows four different SOMs. A SOM is a 2-dimensional visualization spanned by an x-axis and a y-axis.

[0062] The idea behind SOM is to create a grid of X * Y neurons N . Each neuron N contains a weight vector whose length is identical to the length of a feature vector. Determining the number of neurons (X * Y) is a compromise between the amount of data / computational effort on the one hand and the desired reliability of the anomaly detector on the other. In tests already conducted, 20 x 20 = 400 neurons have produced very good results.

[0063] A parameter X is plotted on the x-axis, and a parameter Y on the y-axis, where X and Y define the size of the SOM map. The product X*Y gives the number of neurons represented as honeycombs in the SOM. In the left column with "X=20" there are two SOM maps shown with X=Y=0, 19, and in the right column with "X=25" there are two SOM maps shown with X=Y=0, 24. This results in N=20*20=400 neuron honeycombs in the left column and N=25*25=625 neuron honeycombs in the right column. The two columns are divided horizontally into two rows: the upper row with "AR=64" shows the SOM maps where an autoregression AR with Lag=64 was used to generate the feature vectors, i.e. 64 feature vector elements per feature vector; the lower row with "AR=512" shows the SOM maps where an autoregression AR with Lag=512 was used to generate the feature vectors, i.e. 512 feature vector elements per feature vector.The SOM maps show the unoccupied neuron cells or those with "good" or "error" cells; a "double occupancy" of a neuron cell is also possible. The darker a neuron cell, the greater the "distance" to the neighboring neuron cell. The optimal situation is when as many neuron cells as possible are occupied and there is as little "double occupancy" as possible.

[0064] Learning the SOM is an iterative process. Initially, the weight vectors contain random values. The learning process then proceeds by assigning similar training vectors from the training dataset to the same neuron. "Similar" is to be understood in the sense of a vectorial distance, i.e., depending on the application, the Euclidean distance, the cosine distance, or the Manhattan distance can be used.

[0065] Since the number v of generated feature vectors Fv is usually significantly larger than the number of neurons, one neuron represents many feature vectors. During training, the weight vector G of each neuron is adjusted such that the distance of the weight vector G to the training vectors Fv assigned to the neuron is minimal. In an optimally trained SOM, each neuron has been assigned training vectors. The trained SOM then represents a compact representation of the training data distribution. To use the trained SOM as an anomaly detector, a function is required that assigns a degree to which a feature vector Fv represents an anomaly. For this purpose, the "quantization error" Ev is calculated for each feature vector Fv:

[0066] Ev = | Fv - Gv | for all v, i.e., for all feature vectors. Where | Fv - Gv | is the previously defined vectorial distance, and Gv is the respective weight vector of the "optimal" neuron, i.e., the neuron whose weight vector G has the minimum distance to the feature vector. The quantization error E describes how well the respective feature vector can be mapped to the distribution of the data with which the SOM was trained. A high error indicates that the feature vector is very different from the training data and therefore describes anomalous behavior of the engine.

[0067] Across all feature vectors in the training dataset, the minimum error

[0068] E min = min{Ev} and the maximum error

[0069] E max = max{Ev} .

[0070] During the test phase following the learning phase, in which the device acts as an anomaly detector, the following steps are carried out, similar to the recording phase:

[0071] - There will be again for the time T A Samples plotted on .

[0072] - It is checked whether the engine's load condition was "stationary" during this time, ie, whether there were no significant load changes, and whether the slip was present during this recording in the training phase. A suitable tolerance for the slip can be defined, since the slip determination cannot be carried out error-free, or it must be checked during the training phase which "slip error" would lead to an "anomaly".

[0073] - After successful testing, the feature vector F is generated from the sample values x calculated .

[0074] - For this feature vector, the SOM provides the corresponding neuron and the weight vector G corresponding to the neuron x .

[0075] - From these values ​​the quantization error E x = | F x

[0076] - G x | calculated .

[0077] The simplest variant of an anomaly detector could be realized as follows: A fixed threshold value E T hreshoid is defined for the quantization error. If this is exceeded, there is a high probability that an anomaly has occurred. It is obvious that E Thr eshoid = E max , i.e., the maximum quantization error during the training phase is used as the threshold. Therefore, if the distance between the test vector and the corresponding weight vector becomes too large, an anomaly is present.

[0078] However, this simple variant has disadvantages that will be explained using results from experiments with real engine data. For this purpose, an error-free system (engine) is trained with 100 feature vectors ("good"). Various load conditions are trained. In the test phase, these 100 feature vectors are tested in the "good" state. The system is then intentionally "damaged", e.g. the engine is misaligned ("anomaly"). The test is only carried out under load conditions that were previously trained. A total of another 200 feature vectors are tested.

Claims

Patent claims 1. Method for the condition diagnosis of an electric motor (M) to which an AC supply voltage (U) with a supply frequency (fu) is applied, comprising the following steps: a) recording a time series (I t ) of current values ​​of a machine current (I) which flows through the electric motor (M) due to the supply voltage (U) applied to the electric motor (M); b) describing the time series (I t ) with an autoregression model of order p; c) forming a feature vector (F x ) with p vector elements, where the p vector elements are formed by the p parameters a± of the autoregression model; and d) using the formed feature vector (Fx) in an ML algorithm as training data or as test data.

2. Method according to claim 1, wherein said electric motor (M) is an asynchronous motor.

3. Method according to one of the preceding claims, wherein the time series (I t) is recorded in a phase-locked manner.

4. Method according to one of the preceding claims with the following steps: - Using at least a first feature vector (Fx) as training data of the ML algorithm; - Using at least a second feature vector (Fx) as test data of the ML algorithm; - Making a condition diagnosis by the ML algorithm based on a comparison of the at least one first feature vector (Fx) with the at least one second feature vector (Fx) .

5. Method according to claim 4, comprising the following steps: - detecting, to which at least one first feature vector (Fx) is used as training data, the phase angle at the beginning of the time series (It); - detecting, for at least a second feature vector (Ex) to be used as test data, the phase angle at the beginning of the time series (It); - Using the at least one second feature vector (Ex) as test data of the ML algorithm only if the phase angle assigned to the second feature vector (Ex) at the beginning of the time series (I t ) with a phase angle assigned to a first feature vector (Ex) at the beginning of the time series (I t ) matches .

Images