Method for generating a self-organizing map

Abstract

The invention relates to a method for generating a self-organizing map (SOM) for condition monitoring of an asynchronous motor (M) to which a supply voltage U is applied, wherein a plurality of training vectors are formed to generate the SOM, in each case from a time series I(1...T) of current values It of a current I which, due to the supply voltage U applied to the asynchronous motor (M), flows through the asynchronous motor (M) while the asynchronous motor (M) is in a good condition.

Description

[0001]Description Method for generating a self-organizing map The invention relates to a method for generating a self-organizing map (= SOM) for monitoring the condition of an asynchronous motor. The invention relates to a method, preferably a computer-implemented method, and a device for monitoring the condition of an asynchronous motor. Due to their cost-effective and robust design, asynchronous motors are very widespread in industry: they are 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. Due to the widespread use of asynchronous motors, there is enormous potential in terms of availability and reliability and the associated savings in time and money. The terms “condition monitoring” (= CM, dt.A well-known method for condition monitoring of electric motors is the so-called Motor Current Signature Analysis (MCSA), see, for example, EP3961230A1 (Siemens AG) March 2, 2022. The idea behind it is to measure the currents in the supply cables of an electric motor or the electrical voltage at a connection point. These measured values ​​are then analyzed. For example, the Fast Fourier Transformation (FFT) is used to detect or quantify faults and operating states of the motor in the frequency domain. The classic MCSA is applied in the quasi-stationary case, i.e., at a nominally constant speed. An extension of MCSA is the analysis with short-time FFTs or wavelets in order to analyze processes in a time-resolved manner.Asynchronous motors operate on the principle of electromagnetic induction, meaning interactions between currents or magnetic fields in the stator and rotor generate a torque that drives the rotor shaft. Motor faults such as eccentricity, particularly air gap eccentricity (when the rotational axes of the stator and rotor of the motor do not exactly coincide), broken rotor bars, defective bearings, or misalignment, change or modulate the amplitudes at frequencies characteristic of the faults at a constant slip. Based on the change in the amplitudes of these characteristic frequencies in the frequency spectrum, motor faults can be detected using MCSA. The fault frequencies, in turn, are slip-dependent, meaning a change in slip changes the frequency position of the fault frequencies. Until now, high-quality MCSA required a high frequency resolution.However, this requires long recording times, which leads to large feature vectors and quickly to the "curse of dimensionality." Furthermore, for high-quality MCSA, the monitored system must remain quasi-stationary, which is hardly possible in reality. One object of the present invention is improved condition monitoring of an asynchronous motor. This object is achieved according to the invention by a method having the features specified in claim 1. According to the method for generating a SOM for condition monitoring of an asynchronous motor, several training vectors for generating the SOM are each derived from a time series I(1...T) of current values ​​I. ta current I, which flows through the asynchronous motor due to the supply voltage U applied to the asynchronous motor, while the asynchronous motor (M) is in a good condition. A training vector is a feature vector that is used to train the anomaly detector. The method according to the invention serves to generate a SOM for condition monitoring of an asynchronous motor. The asynchronous motor is operated on an electrical power grid, which supplies the asynchronous motor with a supply voltage. The electrical power grid can be a single- or multi-phase power grid, designed as an island grid or as part of an interconnected grid. In addition to the electric motor in question, the power grid can supply one or more other electrical loads with electrical energy. The voltage in the power grid can be provided by one or more generators and / or one or more batteries. Self-organizing maps (EnglishSelf-organizing maps (SOMs), like most artificial neural networks, operate in two modes: training and mapping. In training mode, an input data set (input space), here the training vectors, is used to generate a map (map space) as a lower-dimensional representation of the input data. In mapping mode, additional input data is classified using the generated map. 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 preferably hexagonal or rectangular grid with preferably 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 position, the weight vectors are approximated to the input data, i.e., the distance, e.g., the Euclidean distance, between the weight vector and the input data is minimized without destroying the topology generated by the map space. After the training phase is completed, 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. A time series y is a sequence of T observation values ​​y. twith t = 1, 2,..., T and a natural order such that the values ​​were observed in the order y1, y2,... The mostly discrete variable t is time, where the values ​​t represent points in time or time intervals. The above-defined object of the invention is also achieved according to the invention by a method having the features specified in claim 3. The inventive method is used for condition monitoring of an asynchronous motor and uses a SOM generated as described above. The inventive method for condition monitoring has a step in which a time series I(1...T) of T current values ​​of a current I, which flows through the asynchronous motor due to the supply voltage applied to the asynchronous motor, is recorded. The inventive method for condition monitoring also has a step in which a slip value s iof the slip of the asynchronous motor is determined, which the asynchronous motor exhibits during the acquisition of said time series I(1...T). The inventive method for condition monitoring also comprises a step in which a feature vector Fx is generated from the current values ​​of the acquired time series I(1...T) if the determined slip value s i , with a given tolerance, was already present when generating one of the training vectors Fν of the SOM. The inventive method for condition monitoring also comprises a step in which a neuron N xy the said SOM and a neuron N xy associated weight vector G xyis assigned. The inventive method for condition monitoring also comprises a step in which a quantization error Ex for the feature vector Fx is calculated as Ex = |Fx – Gx|. The inventive method for condition monitoring also comprises a step in which an anomaly of the asynchronous motor is determined if the quantization error Ex exceeds a fixed threshold value E Threshold exceeds: Ex > E ThresholdThe inventive method for condition monitoring also has a step in which the anomaly is reported if an anomaly is detected. The invention relates to an anomaly detection method based on an unsupervised learning method. In a first part of a training phase, a so-called recording phase, an asynchronous motor is operated for a limited period of time in a "good state", i.e., in error-free operation. During this time, measured values, e.g., measured values ​​of the current flowing through the motor windings, are recorded, and training vectors are generated therefrom, which describe this "good state". In a second part of the training phase, a so-called learning phase, an anomaly detector (= AD) is trained using several of these training vectors. This means that the AD only sees the "good state" of the asynchronous motor, but does not know the "bad state", i.e.faulty operation of the asynchronous motor. After the training phase has ended, a test phase, also known as the detection phase, of the anomaly detector is activated. This means that from this point on, the anomaly detector monitors the asynchronous motor for operating states that deviate from the learned "good state" of the asynchronous motor and reports these to an operator. In detail, the following phases are run through one after the other: 1. Training phase 1.1. Recording phase (generation of training vectors of the "good state") 1.2. Learning phase (learning the "good state" of the asynchronous motor) 2. Test phase (anomaly detection) The "Self-Organizing Map" method, also known as the "Kohonen map" after its inventor Prof. Teuvo Kohonen (1934 - 2021), has proven to be a particularly suitable basis for such anomaly detection for asynchronous motors. This method can be expanded into an anomaly detector through modifications.This requires input values, the so-called "training vectors," which are calculated from the input data. As described in the theory of MCSA, the frequency spectrum of the motor current of an asynchronous motor—and thus naturally also the time signal of the motor current—is highly dependent on the load condition of the asynchronous motor, i.e., on the slip s of the asynchronous motor. Depending on the number of pole pairs and the number of rotor bars, different slip values ​​si lead to different modulation artifacts in the spectrum. The following basic assumptions are made: • s. min indicates the minimum slip (= at idle). • s Nrefers to the slip at nominal load. • Training and test phases take place with the same or an identical asynchronous motor: this ensures that the asynchronous motor does not change after the training phase - if it does, the anomaly detector must be re-taught. So far, only MCSA solutions in conjunction with supervised learning are known. In addition, the "classic" MCSA uses an FFT. The resulting disadvantages and problems are solved by the present invention: - Supervised learning has the disadvantage that training must be carried out with both "good" and "bad" data. In a customer system, recording "bad" data, i.e. error states, for training is usually not possible because this would require intentionally damaging a motor. If the data recording were to take place at the manufacturer's site, the manufacturer would have to supply data for a large number of different motors.In addition, the data recorded by the manufacturer would not take the customer's operating environment into account, which could lead to erroneous detection. An anomaly detector such as that in the present invention, on the other hand, can be trained "unsupervised," i.e., exclusively with "good" data. - "Classic" MCSA, i.e., a frequency analysis, uses discrete frequencies that result from the physical properties of the motor. Errors such as misalignment only change the physical properties of the motor to a very small extent ("eccentricity"), so that such an error is very difficult to detect. A time-domain analysis such as that in the present invention, on the other hand, can also reliably detect such errors. - For high-quality "classic" MCSA, i.e., a frequency analysis, a high frequency resolution is required. However, this requires long recording times, during which the system must remain stationary.In addition, very high-dimensional feature vectors result, which would impair the performance of the anomaly detector due to the "curse of dimensionality." These problems are circumvented with a time-domain analysis as in the present invention. Advantageous embodiments and further developments of the invention are specified in the dependent claims. According to a preferred embodiment of the method for generating a SOM, a time series I(1...T) of T current values ​​I is initially generated in a first step. t of a current I which flows through the asynchronous motor due to the supply voltage U applied to the asynchronous motor, while the asynchronous motor is in a good condition and the slip s of the asynchronous motor is within a specified slip tolerance band (-Δs lo ; +Δs hi ) by a slip value s i The slip s therefore lies in the value interval [s i - Δs lo ;s i + Δshi ]. Then, in a second step, a training vector Fν with T vector elements is formed, where the T vector elements are determined by the T current values ​​I t the recorded time series I(1...T) and the training vector Fν is assigned the slip value s i The two steps are repeated one or more times with different slack values ​​to obtain a group of training vectors as a training data set. A matrix of X times Y neurons N xy generated, where each of the neurons N xy each a weight vector G xy of length T. A neuron N xy those training vectors Fν,xy from the training data set are assigned whose vectorial distance δ(Fν;N xy ) to the weight vector G xy of the neuron N xy does not exceed a given distance limit δmax. And the weight vectors G xy of the neurons Nxy are adjusted in such a way that the vectorial distance δ(G;Fν,xy) of a neuron N xy assigned weight vector G xy to the neuron N xyassigned training vectors Fν,xy is minimal. According to a preferred embodiment of the method for monitoring the condition of an asynchronous motor, the analysis and detection of an anomaly is carried out by at least one trained AI function of the anomaly detector. The method therefore uses machine methods to detect an anomaly. The current variations that occur in reality can be very diverse because there are practically an infinite number of different possibilities for the structure and elements of a power grid and a drive train. It can therefore be difficult to define all the criteria for when an anomaly exists in advance in such a comprehensive manner that it can be clearly determined for each possible current variation whether an anomaly exists or not.An AI function is advantageous in that it can continuously learn and evolve. This means that the AI ​​function can apply learned criteria to current variations that occur in reality and that an operator had not previously considered, allowing it to classify whether an anomaly is present or not. By applying the trained artificial intelligence function, or AI function for short, to the measured time series of current values, function output data is generated that contains information about anomalies in an asynchronous motor. The trained anomaly detection function is capable of extracting information regarding the quality of the operating condition of an asynchronous motor from the measured time series of current values. The trained function is a function trained by an ML algorithm (ML = Machine Learning).Training is generally understood to mean the optimization of a mapping of input parameters of a parameterized system model, e.g., a neural network, to one or more target parameters. This mapping is optimized during a training phase according to predetermined, learned and / or to-be-learned criteria. For pattern recognition models, for example, criteria can be a percentage of successfully recognized patterns. A training structure can, for example, comprise a network structure of neurons in a neural network and / or weights of connections between the neurons, which are formed during training so that the criteria are met as well as possible. In the present exemplary embodiment, a neural network, abbreviated to NN, is trained to detect an anomaly in an asynchronous motor using predetermined temporal current profiles.In order to distinguish a specific current waveform pattern from other current waveform patterns, the properties of the current waveform pattern must be described and mathematically mapped. The more precisely the pattern can be described and the more analyzable information is available, the more reliably the pattern recognition works. To measure motor quality, current waveform data sets are created and used to train an anomaly detector. The trained function is part of a pattern recognition device. For configuration, the pattern recognition device has one or more trainable functions, i.e. algorithms that can be executed in the pattern recognition device's computing modules and implement machine learning methods to optimize pattern recognition through training.These trainable functions can be trained using known standard machine learning methods to recognize temporal current curves specified in input files as accurately as possible. As a rule, training is more successful the more training data, i.e., files that can be used for training, are available. In particular, the training data should representatively cover the largest possible range of possible configurations of the asynchronous motor. This includes, in particular, training data for different slip values ​​of the asynchronous motor, i.e., under different load conditions. As already mentioned above, the aim of the present invention is to train the neural network NN so that it can recognize a motor anomaly. An SOM algorithm is used to train the neural network NN.In general, a trained function imitates cognitive abilities that people associate with the minds of other people. In particular, through training on the basis of training data, a trained function is able to adapt to new circumstances and discover and extrapolate patterns. In general, parameters of a trained function can be adjusted through training, i.e., learning. In general, supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning can be used for this purpose. In addition, representation learning, also called “feature learning,” can be used. In particular, the parameters of the trained functions can be adjusted iteratively through multiple training steps. The present invention, in contrast, uses unsupervised learning.According to a preferred embodiment of the invention, feature data is determined from the measured current values, which are used to train the anomaly detector. Determining feature data, e.g., a mean value, minimum and / or maximum values, average values, expected value, and variance of a distribution, from measured current measurement series is generally accompanied by a reduction in the amount of data, which significantly simplifies and accelerates data processing: instead of the extensive measured current measurement series, the anomaly detector then only has to calculate with the feature data. According to a preferred embodiment of the invention, E. Th- reshold = E max chosen, where E maxis the maximum of all quantization errors Eν of the group of training vectors Fν used as a training data set. The "anomaly limit" is thus calculated as the maximum deviation of the "good" data. A measurement whose quantization error Ex exceeds the value E max is then classified as an anomaly. According to a preferred embodiment of the invention, chosen, where E min the minimum and E max the maximum of all quantization errors E ν the group of training vectors F ν which are used as a training dataset, L is set as a sensitivity level L ∈[-1, … ,1], and where percentile(E ν,Train ,L) specifies the value for which the anomaly score for (100 * L)% of the training examples is smaller than this value. The following cases are possible: Case 1: L = 0. This value leads to a threshold value E Threshold = E max. The maximum error during training defines the threshold for anomaly detection (as described above). Case 2: L > 0. The system should be more "sensitive." For L = 1, E Threshold = E min , meaning the system would classify every training vector in the training data set as an anomaly. The system would therefore be maximally sensitive. In general, the sensitivity can be set between 0% and 100%. Case 3: L < 0. The system should be "less sensitive." For this purpose, a value between -1 and 0 is set, so that a threshold value E Threshold > E max can be achieved. A value of L = -1 leads to a threshold value of 2*Emax - Emin. In this case, the histogram is mirrored at the point E = Emax. After the optimal threshold value E has been determined from the desired sensitivity level L, Thresholdhas been calculated and adjusted, the system can be used for anomaly detection. According to a preferred embodiment of the invention, the parameters N and p are defined such that an anomaly of the asynchronous motor is detected if at least (100 * p)% of the last N time windows have been classified as anomalies by the anomaly detector, and otherwise the normal state of the motor is assumed. According to a preferred embodiment of the invention, a pointer representation is used for monitoring, wherein the quantization error of the feature vector belonging to the current time window is depicted using a pointer that can move along a scale. The pointer moves dynamically with each incoming time window. An additional limit value E Warning < E Thresholddefined: from which an increased anomaly value already exists, but is not yet classified as faulty. This achieves a visualization of the magnitude of the anomaly: the "green range" is defined by the range of the "good data"; a higher anomaly indicates an engine problem, the larger the value, the worse: yellow or red. According to a preferred design of the vector display, L>0 (E Threshold > E max ) and E Warning = E Max This means that an additional limit value E Warning < E Thresholddefined which indicates the error above which an increased anomaly value already exists, but which is not yet classified as faulty. In this way, a warning range can be defined in which a user is warned of an impending anomaly. A further aspect of the invention is a device for monitoring the condition of a drive train which has an asynchronous motor operated on an electrical power grid. The device has sensors for detecting a current resulting from a voltage applied to the motor, which current flows through the asynchronous motor. The device has a first sensor unit for detecting a time series of current values ​​of a current which flows through the asynchronous motor due to the supply voltage applied to the asynchronous motor.The device comprises a second sensor unit for determining the slip value of the slip of the asynchronous motor which the asynchronous motor has during the acquisition of said time series. The device has a computing unit which is adapted to generate a feature vector from the current values ​​of the acquired time series if the determined slip value, with a predetermined tolerance, already existed when generating one of the training vectors according to the method for generating a SOM, to assign to the feature vector a neuron Nx of said SOM and a weight vector Gx assigned to the neuron Nx, which was formed according to the method for generating a SOM, to calculate a quantization error Ex for the feature vector Fν: Ex = |Fx - Gx|, and to detect an anomaly of the asynchronous motor if the quantization error Ex exceeds a fixed predetermined threshold value E. Thresholdexceeds. The device also has an output unit for reporting the anomaly. A further aspect of the invention is a computer program which comprises instructions which, when the program is executed on a computer, cause the device for condition monitoring of an asynchronous motor to carry out the method for condition monitoring of an asynchronous motor. A further aspect of the invention is a computer program which comprises instructions which, when the program is executed on a computer, cause the computer to carry out the inventive method for generating an SOM. The invention is explained below with the aid of the accompanying drawings. They show schematically and not to scale: Fig. 1 a first embodiment of a device according to the invention for condition monitoring of a drive train; Fig.2 shows a second embodiment of a device according to the invention for condition monitoring of a drive train; Fig. 3 is a flow diagram of an embodiment of the method according to the invention for generating an SOM for condition monitoring of an asynchronous motor; Fig. 4 is a flow diagram of an embodiment of the method according to the invention for condition monitoring of an asynchronous motor; Fig. 5 is an embodiment of an SOM; Fig. 6 shows histograms of a quantization error for normal states and error states for a first case; Fig. 7 is a confusion matrix of the anomaly detector; Fig. 8 shows histograms of a quantization error for normal states and error states for a second case; and Fig. 9 is a pointer instrument for dynamic monitoring of anomaly detection. Fig. 1 shows an arrangement with an asynchronous motor M, which is connected via supply lines 11 to a voltage source 10, which has a supply voltage U. supply with a voltage frequency f supplyThe 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 values ​​are recorded at the sampling frequency f A [1 / T] for a duration T A [T] is recorded. 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, resulting in N A =4096 = 2 12 Sample values. The N A Current measurement values ​​are sent from the current sensor 14 to a computing unit 16. From these N A For each current value, the computing unit 16 generates a feature vector of length LF = N Aformed. The method steps for determining the slip s of the asynchronous motor M are also carried out 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 determining the slip s from current measurements. This software is executed by the processor 18. Input values ​​such as the number of neurons can be transferred to the computing unit 16 via 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. Fig. 2 shows an alternative arrangement in which the asynchronous motor M is designed as a drive machine of a drive train D, which also has a mechanical transmission G and a working machine W.A torque provided by the asynchronous motor M is transmitted via the gearbox G to the working machine W. The working machine can be, for example, a conveyor belt, a roller or a cylinder. In addition to the asynchronous motor M of the drive train D, further electrical loads 24 are connected to the power grid N. Fig. 3 shows a flow diagram of an embodiment of the method according to the invention for generating an SOM for monitoring the condition of an asynchronous motor M to which a supply voltage U is applied. In a first step 310, T current values ​​I. t of an electric current I flowing through the electric motor M, which flows through the asynchronous motor M due to the supply voltage U applied to the asynchronous motor M. This gives a time series I(1...T) of T current values ​​I t , while the asynchronous motor M is in good condition. During the recording phase, current values ​​I trecorded, which describe the "good condition" of the motor. These current values ​​I t are then used to train the anomaly detector. The following aspects are very important: - The system should have as many load conditions or slip values ​​as possible i in the slip value range [s min , ..., s N ] For each recording, the respective slip value s i as “label” L s - If a load condition is recorded, the load must be T A the load condition (= slip) should be approximately constant: L s= constant. It is pointless to record the current values ​​during large load changes (= slip changes), because otherwise, different states and thus different frequency spectra would overlap. - The recording phase is completed when a sufficiently large variation of slip values ​​has been recorded, or when the maximum recording time has been reached. Step 320: Generate, from the current values ​​I t the recorded time series, a training vector Fν, which is assigned a slip value s i of the slip s of the asynchronous motor. Step 330: Repeat steps 310 and 320 for different slip values ​​s i, to obtain a group of training vectors {Fν1, Fν2, Fν3, ...} as a training data set. Step 340: Generate a matrix of X times Y neurons, where each of the neurons is assigned a weight vector G of length T. Step 350: Assign to a neuron those training vectors Fν from the training data set whose vectorial distance to the neuron, more precisely: to the weight vector G of the neuron, does not exceed a predetermined distance threshold. Step 360: Adjust the weight vectors G of the neurons such that the distance of the weight vector G to the training vectors (Fν) assigned to a neuron is minimal. The recording phase is followed by the learning phase. During this phase, the recorded neurons and labeled with the slack label L slabeled training vectors are learned into the SOM. Fig. 4 shows a flowchart of an embodiment of the inventive method for condition monitoring of an asynchronous motor M, to which a supply voltage U is applied, using a SOM. Step 410: Acquisition of a time series I(1...T) of T current values ​​of a current I, which flows through the asynchronous motor M due to the supply voltage U applied to the asynchronous motor M. Step 420: Determination of the slip value s i of the slip s of the asynchronous motor M, which the asynchronous motor M exhibits during the acquisition of the said time series I(1...T). Step 430: If the determined slip value s i , with a given slip tolerance band (-Δs lo ; +Δs hi), already present when generating one of the training vectors Fν according to the method for generating the SOM, generating a feature vector Fx from the current values ​​of the acquired time series I(1...T). Step 440: Assigning, to the feature vector Fx, a neuron Nx of said SOM and a weight vector Gx assigned to the neuron Nx. Step 450: Calculating a quantization error Ex for the feature vector Fx as Ex = |Fx - Gx|. Step 460: Determining an anomaly of the asynchronous motor M if the quantization error Ex exceeds a fixed threshold value E Threshold exceeds: Ex > E Threshold. Step 470: If an anomaly is detected, report the anomaly. Fig. 5 shows four different SOMs. An SOM is a 2-dimensional visualization spanned by an x-axis and a y-axis. The idea of ​​the 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 led to very good results. A parameter X is plotted on the x-axis, and a parameter Y on the y-axis, with X and Y defining 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” two SOM maps are shown with X=Y=0, ..., 19, In ​​the right column with "X=25" two SOM maps are 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 top row with "AR=64" shows the SOM maps in which an autoregression AR with Lag=64 was used to generate the feature vectors, i.e. 64 feature vector elements per feature vector; the bottom row with "AR=512" shows the SOM maps in which 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, those with "good" or "defect" cells; a neuron cell can also be "double-occupied." The darker a neuron cell, the greater the "distance" to the neighboring neuron cell.The optimal solution is when as many neuron honeycombs as possible are occupied and, if possible, there is no "double occupancy". Learning the SOM is an iterative process. In the initial state, the weight vectors contain random values. The learning process then occurs such that similar training vectors from the training data set 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. Since the number ν of generated feature vectors Fν 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 Fν assigned to the neuron is minimal. In an optimally trained SOM, training vectors were assigned to each neuron.The trained SOM then represents a compact representation of the training data distribution. In order to use the trained SOM as an anomaly detector, a function is required that assigns a degree to which a feature vector Fν represents an anomaly. For this purpose, the "quantization error" Eν is calculated for each feature vector Fν: Eν = |Fν - Gν| for all ν, i.e., for all feature vectors. Here, |Fν - Gν| is the previously defined vectorial distance, and Gν 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.The minimum error E is now calculated across all feature vectors in the training dataset. min = min{Eν} and the maximum error E max = max{Eν} is determined. During the test phase following the learning phase, in which the device operates as an anomaly detector, the following steps are carried out, similar to the recording phase: - The time T A Samples are recorded. - It is checked whether the engine 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". - After a successful check, the feature vector F is generated from the samples. xcalculated. - For this feature vector, the SOM provides the associated neuron and the weight vector G corresponding to the neuron x . - From these values ​​the quantization error E x = | F x - G x | calculated. The simplest version of an anomaly detector could be implemented as follows: A fixed threshold value E Threshold for the quantization error. If this value is exceeded, there is a high probability that an anomaly has occurred. It is obvious that E Threshold = E maxto be selected, i.e. the maximum quantization error during the training phase is used as the threshold. If the distance between the test vector and the corresponding weight vector is too large, an anomaly exists. However, this simple variant has disadvantages that will be explained using results from experiments with real motor data. For this purpose, an error-free system (motor) 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 motor is misaligned (“anomaly”). The test is only carried out under load conditions that were previously trained. A total of 200 additional feature vectors are tested. Fig. 6 shows the quantization errors of the feature vectors, labeled “error” on the x-axis.The y-axis shows the number of candidates (= scenarios) for each error (frequency). In total, there are 100 "good scenarios" (NGD=100) and 200 "failure scenarios" (NAD=200); therefore, the sum of all bars is 100 + 200 = 300. -E denotes the frequency of the error. ν,Train The errors of the training vectors generated during the training phase. The data are labeled "good" and without hatching. It is clearly visible that the errors are subject to an unknown distribution, but the maximum error is approximately 0.16, which also corresponds to the detection threshold. - E ν,Testthe errors of the feature vectors generated during the test phase. The data is labeled "anomaly" and hatched. Here, too, it is clear that the errors are subject to an unknown distribution, but all errors are > 0.6, so that the anomaly is clearly detected for each state. The histogram represents the number of candidates / scenarios for the respective anomaly. Fig. 7 shows that a correspondingly good truth matrix, also known as a confusion matrix, results for the anomaly detector. All 100 training states for the "good" motor are recognized as "good" (or "healthy"), and the 200 test states are also correctly assigned. Fig. 8 shows a histogram representation analogous to that in Fig. 6 for an application in which an error leads to only very small changes in the current signal. The system was trained and tested with 900 “good” vectors and tested with a further 1800 “anomaly” vectors.Since the "good" and "anomaly" histograms are not completely separable, two possible types of misclassification by the anomaly detector arise: - False Positive: The anomaly detector identifies a normal state as faulty. - False Negative: The anomaly detector identifies an error state as normal. Both error types can be of varying importance in different applications. In certain applications, it is necessary to generate as few false alarms (false positives) as possible, since, for example, every alarm would lead to a manual inspection of the machine and would generate correspondingly high costs. In other applications, it is more important to have the earliest possible indication of a problem in order to be able to plan for a failure as far in advance as possible. To ensure the configurability of the anomaly detector with regard to reliability, a fixed threshold for the detection of an anomaly is first determined.A reasonable threshold should be within a certain interval around E. max A smaller value leads to a higher sensitivity of the anomaly detector, meaning that more alarms and thus potentially more false alarms are produced. A higher threshold leads to fewer alarms and thus potentially fewer false alarms, but also a potentially lower detection rate. The errors during the learning phase are between E min = min{ Ev} and E max = max{ Ev} with the fixed distribution Ev, Train, which is visualized by the histogram. To determine the threshold, the system user first defines a "sensitivity level" L ∈ [-1, … , 1]. The threshold is then calculated as where returns the value for which The anomaly score for (100 * L) % of the training examples is smaller than this value. The following cases are possible: Case 1: L = 0. This value leads to a threshold value E Threshold = E max . The maximum error during training defines the threshold for anomaly detection (as described above). Case 2: L > 0. The system should be more "sensitive." For L = 1, E Threshold = E min , meaning the system would classify every feature vector in the training data set as an anomaly. The system would therefore be maximally sensitive. In general, the sensitivity can be set between 0% and 100%. Case 3: L < 0. The system should be "less sensitive." For this purpose, a value between -1 and 0 is set, so that a threshold value E Threshold > E maxcan be achieved. A value of L = -1 leads to a threshold value of 2*Emax - Emin. In this case, the histogram is mirrored at the point E = Emax. After the optimal threshold value E has been determined from the desired sensitivity level L, Threshold has been calculated and adjusted, the system can be used for anomaly detection. To further reduce the false alarm rate, alarms for isolated anomalies are suppressed. For this purpose, two additional parameters N and p are defined. If at least (100 * p)% of the last N time windows have been classified as anomalies by the anomaly detector, an alarm is issued. Otherwise, the normal state of the engine is assumed. For example, with a recording time of T A= 1.28 seconds and values ​​N = 10 and p = 50, an alarm is only triggered if at least half of all time windows have been classified as anomalies during the last 12.8 seconds. Fig. 9 shows a "pointer instrument" that can be used to monitor the system. The "pointer instrument" can be used on a monitor, on the web, or on a mobile phone. On axis 81, the pointer 80 displays the quantization error of the feature vector associated with the current time window. The pointer 80 moves dynamically with each incoming time window. An additional limit value E Warning < E Threshold which indicates the error at which an increased anomaly value is already present, but which is not yet classified as faulty. For example, L>0 (E Threshold > E Max ) and E Warning = E MaxThe left area 82 (“green area”) would then correspond to the error range of the training data. As soon as a time window with an error greater than E Max occurs, the pointer 80 moves to the middle area 83 (“warning area”). The pointer 80 moves to the right area 84 (“red area”) as soon as the threshold value E Threshold is exceeded. If E Threshold If the limit is exceeded sufficiently often (see above), an additional alarm is issued.

Claims

Claims 1. Method for generating a self-organizing map (= SOM) for monitoring the condition of an asynchronous motor (M) to which a supply voltage U is applied, wherein several training vectors for generating the SOM are each derived from a time series I(1...T) of current values ​​I t a 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, are formed.

2. Method according to claim 1, comprising the following steps: a) Obtaining a time series I(1...T) of T current values ​​I t a 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 condition and the slip (s) of the asynchronous motor is within a specified slip tolerance band (-Δs lo ; +Δs hi ) by a slip value s ib) Forming a training vector Fν with T vector elements, where the T vector elements are determined by the T current values ​​I t of the recorded time series I(1...T) and the training vector Fν the slip value s i is assigned; c) Repeating steps a) and b) once or several times with different slip values ​​(s i1 , s i2 , s i3 , ...), to obtain a group of training vectors (Fν1, Fν2, Fν3, ...) as a training data set; d) generating a matrix of X times Y neurons N xy , where each of the neurons N xy each a weight vector G xy of length T; e) Assign to a neuron N xy , those training vectors Fν,xy from the training data set (Fν1, Fν2, Fν3, ...) whose vectorial distance δ(Fν;N xy ) to the weight vector G xy of the neuron N xydoes not exceed a predetermined distance limit δmax; and f) adjusting the weight vectors G xy of the neurons N xy such that the vectorial distance δ(G;Fν,xy) of a Neuron N xy associated weight vector G xy to the neuron N xy associated training vectors Fν,xy is minimal.

3. Method for monitoring the condition of an asynchronous motor (M) to which a supply voltage (U) is applied, using a SOM generated according to a method according to one of the preceding claims, comprising the following steps: a) detecting a time series I(1...T) of T current values ​​of a current I flowing through the asynchronous motor (M) due to the supply voltage (U) applied to the asynchronous motor (M); b) determining a slip value s ithe slip (s) of the asynchronous motor (M) exhibited by the asynchronous motor (M) during the acquisition of said time series I(1...T); c) If the determined slip value s i , with a given tolerance (-Δs1; +Δs2), already present when generating one of the training vectors Fν of the SOM, generating a feature vector Fx from the current values ​​of the acquired time series I(1...T); d) Assigning, to the feature vector Fx, a neuron N xy the said SOM and one of the said neuron N xy associated weight vector G xy ; e) Calculating a quantization error Ex for the feature vector Fx as Ex = |Fx – Gx|. f) Detecting an anomaly of the asynchronous motor (M) if the quantization error Ex exceeds a fixed threshold value E Threshold exceeds: Ex > E Threshold; e) upon detection of an anomaly, reporting the anomaly.

4. The method according to claim 3, wherein the analysis and detection of an anomaly is carried out by an anomaly detector (16) with at least one trained AI function (18).

5. The method according to one of claims 3 and 4, wherein E Threshold = E max is chosen, where E max is the maximum of all quantization errors Eν of the group of training vectors (Fν1, Fν2, Fν3, ...) used as a training data set.

6. Method according to one of claims 3 to 5, wherein is chosen, where E min the minimum and E max is the maximum of all quantization errors Eν of the group of training vectors (Fν1, Fν2, Fν3, ...) used as a training data set, L is defined as a sensitivity level L ∈[-1, … ,1], and where percentile(E ν,Train,L) is the value for which the anomaly score for (100 * L)% of the training vectors is smaller than this value.

7. The method according to one of claims 3 to 6, wherein the parameters N and p are defined such that an anomaly of the asynchronous motor is detected if at least 100 * p% of the last N time windows have been classified as anomalies by the anomaly detector, and otherwise the normal state of the motor is assumed.

8. The method according to one of claims 3 to 7, wherein a pointer representation is used for monitoring, wherein the quantization error of the feature vector belonging to the current time window is depicted on the pointer scale with a pointer, wherein the pointer moves dynamically with each incoming time window, and wherein an additional limit value E Warning < E Thresholdwhich indicates the error at which an increased anomaly value already exists, but which is not yet classified as faulty.

9. Method according to claim 8, wherein L>0 and E Warning = E Max 10. Device for monitoring the condition of an asynchronous motor (M) to which a supply voltage (U) is applied, using a SOM generated according to a method according to one of claims 1 and 2, comprising: a first sensor unit for detecting a time series (I t ) of T current values ​​of a current (I) which flows through the asynchronous motor (M) due to the supply voltage (U) applied to the asynchronous motor (M); a second sensor unit for determining the slip value of the slip (s) of the asynchronous motor (M) which the asynchronous motor (M) has during the recording of said time series (I t); a computing unit adapted to - generate a feature vector (Fx) from the current values ​​of the acquired time series (I t ), if the determined slip value, with a predetermined tolerance, already existed when generating one of the training vectors (Fν), - Assigning a neuron Nx of said SOM and a weight vector Gx assigned to the neuron Nx to the feature vector (Fx); - Calculating a quantization error Ex for the feature vector Fν: Ex = |Fx – Gx|. - Determining an anomaly of the asynchronous motor (M) if the quantization error Ex exceeds a fixed threshold value E Threshold exceeds; and an output unit for reporting the anomaly.

11. A computer program comprising instructions which, when executed on a computer, cause the apparatus of claim 10 to perform the method of any one of claims 3 to 9.

Images