Method for generating self-organizing mappings

Abstract

The invention relates to a method for generating a self-organizing map (= SOM) for monitoring the state of an asynchronous motor (M) to which a supply voltage U is applied, a plurality of training vectors for generating the SOM being formed in each case from a time sequence I (1... T) of the current value It of the current I, the current flows through the asynchronous motor (M) as a result of a supply voltage U applied to the asynchronous motor (M), during which the asynchronous motor (M) is in a good state.

Description

TECHNICAL FIELD

[0001] The present invention relates to a method for generating a self-organizing map (= SOM) for condition monitoring of an asynchronous electric motor. The invention relates to a method, preferably a computer-implemented method, for condition monitoring of an asynchronous electric motor, and to a device for condition monitoring of an asynchronous electric motor. BACKGROUND

[0002] Asynchronous electric motors are very popular in industry due to their cost-effective and robust design: they are found in many applications, such as fans, saws and pumps, from relatively small devices in the watt range to devices in the megawatt range. As a result, asynchronous electric motors belong to the largest consumers of global power generation. Due to the widespread use of asynchronous electric motors, there is a great potential for availability and reliability, and consequently for saving time and money. In this context, the terms "condition monitoring" (= CM) and "predictive maintenance" (= PM) have emerged.

[0003] A known method for condition monitoring of electric motors is the so-called motor current signature analysis (= MCSA); see, for example, EP 3961230 A1 (Siemens AG) of 2 March 2022. The idea behind this is to measure the current in the motor supply cable and the voltage at the connection points. These measured values are then analyzed. For example, a fast Fourier transformation (= FFT) is used to detect and quantify the fault and operating state of the electric motor in the frequency domain. Here, the conventional MCSA is applied to quasi-stationary cases, i.e. nominally constant rotational speed. An extension of the MCSA is the use of short-term FFTs or wavelets for analysis in order to analyze the process in a time-resolved manner.

[0004] Asynchronous electric motors operate on the principle of electromagnetic induction, i.e. the interaction between currents and magnetic fields in the stator and rotor produces a torque, which drives the rotor shaft. Motor faults, such as eccentricity, in particular air gap eccentricity (when the rotational axes of the stator and rotor of the electric motor do not completely coincide), broken bars, bearing damage or misalignment, change or modulate the amplitude of fault characteristic frequencies at constant slip. As a result, motor faults can be identified by means of MCSA based on the amplitude changes of these characteristic frequencies in the frequency spectrum. The fault frequencies themselves are dependent on the slip, i.e. a change in the slip changes the frequency position of the fault frequencies.

[0005] So far high quality MCSA required high frequency resolution. But this requires long recording times, which leads to an excessive feature vector and quickly to the "Fluch der Dimensionalität". Furthermore, for high quality MCSA the monitored system has to stay quasi-stationary, but this is hardly achievable in reality. SUMMARY

[0006] The technical problem to be solved by the present invention is to improve the condition monitoring of asynchronous motors.

[0007] According to the invention, the above-mentioned technical problem is solved by a method having the features of claim 1. According to the method for generating a SOM for condition monitoring of an asynchronous motor, a time series I (1...T) of current values I of a current I flowing through the asynchronous motor due to a supply voltage U applied to the asynchronous motor is formed, during which the asynchronous motor (M) is in a good condition. The training vectors are feature vectors for training the anomaly detector. t

[0008] The method according to the invention is for generating a SOM for condition monitoring of an asynchronous motor. The asynchronous motor is operated on an electrical grid, which provides a supply voltage to the asynchronous motor. The electrical grid can be a single- or multi-phase electrical grid, which is designed as an island network or as part of an interconnected network. In addition to the motor under consideration, the electrical grid can also provide electrical energy for one or more other electrical loads. The voltage in the electrical grid can be provided by one or more electrical generators and / or one or more batteries.

[0009] Like most artificial neural networks, a self-organizing map (SOM) works in two modes: training and mapping. In the training mode, an input data set ("input space") is used, here the training vectors, in order to generate a map ("map space") as a low-dimensional representation of the input data. In the mapping mode, additional input data is classified with the help of the generated map.

[0010] The goal of the training is to represent the "input space" of p dimensions (i.e. with p variables) as a "map space" of 2 dimensions. The "map space" consists of components, which are called nodes or neurons and which are arranged on a preferably hexagonal or rectangular grid with preferably two dimensions. The number of nodes and their position on the grid are predetermined based on the goal to be tracked based on an analysis and examination of the "input space".

[0011] ​A weight vector is assigned to each node in the "mapped space" that defines the node's position in the "input space". When the node is anchored at its position in the "mapped space", the weight vector approximates the input data, i.e., the distance (e.g., Euclidean distance) between the weight vector and the input data is minimized without destroying the topology generated by the "mapped space".

[0012] After 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 a time series of current values, by identifying the node whose weight vector is closest to the "input space" vector, i.e. with the smallest distance metric (e.g. Euclidean distance). A time series y is a sequence of T observations y t , where t = 1, 2, ..., T, and has a natural ordering such that values ​​are observed in the sequence y1, y2,.... Usually the discrete variable t is time, and the values ​​t represent time points or time intervals.

[0013] The above technical problem is also solved according to the present invention by a method having the features described in claim 3. The method according to the present invention is used for state monitoring of an asynchronous motor and uses the SOM generated as described above. The method according to the present invention for state monitoring comprises the following steps, wherein a time series I (1...T) of T current values ​​of a current I flowing through the asynchronous motor due to the supply voltage applied to the asynchronous motor is detected. The method according to the present invention for state monitoring also comprises the following steps, wherein a slip value s of the asynchronous motor, which the asynchronous motor has during the detection of the time series I (1...T), is determined. i The method for condition monitoring according to the present invention further comprises the following step, wherein if the determined slip value s i By pre-determined tolerance, the training vectors for SOM are generated. If one of them already exists, a feature vector Fx is generated from the current value of the detected time series I(1...T). The method for state monitoring according to the present invention also includes the following steps, wherein the neurons N of the SOM are xy and assigned to the neuron N xy The weight vector G xy Assigned to the feature vector Fx. The method for condition monitoring according to the invention further comprises the following step, wherein the quantization error Ex for the feature vector Fx is calculated, The method for state monitoring according to the present invention further comprises the following step, wherein if the quantization error Ex exceeds a fixed predetermined threshold value E Threshold : an anomaly of the asynchronous motor is determined. The method for performing a condition monitoring according to the present application further comprises the step, wherein if an anomaly is determined, the anomaly is reported.

[0014] The present application relates to anomaly detection based on unsupervised learning methods. In this case, the asynchronous motor is operated in a first part of a training phase, called recording phase, in a "good condition", i.e. without fault, for a limited period of time, wherein during this time, measurement values, for example of the current strength of the current flowing through the motor windings, are recorded and from which a training vector is generated which describes this "good condition". In a second part of the training phase, called learning phase, a plurality of these training vectors are used to train an anomaly detector (= AD), i.e. the AD only sees the "good condition" of the asynchronous motor, but not the "bad condition", i.e. the fault operation, of the asynchronous motor. After the training phase, a test phase, also called detection phase, of the anomaly detector is activated, i.e. from this point in time, the anomaly detector monitors the asynchronous motor in terms of operating states deviating from the learned "good condition" of the asynchronous motor and reports this to the operator. The following phases are therefore run in particular in this order:

[0015] 1. Training phase

[0016] 1.1. Recording phase (generation of training vectors of "good condition")

[0017] 1.2. Learning phase (learning of "good condition" of asynchronous motor)

[0018] 2. Test phase (anomaly detection)

[0019] It has proven particularly suitable as a basis for such anomaly detection for asynchronous motors that the "self-organizing map" method, also known as "Kohonen map", named after its inventor Professor Teuvo Kohonen (1934 - 2021), can be extended to an anomaly detector by modification. For this purpose, input values, so-called "training vectors", are required, which are calculated from input data.

[0020] As described in the theory of MCSA, the spectrum of the motor current of an asynchronous motor, and thus of course also the time signal of the motor current, is highly dependent on the load state of the asynchronous motor, i.e. 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:

[0021] • s min denotes the minimum slip (= at no-load operation).

[0022] • s N denotes the slip at rated load.

[0023] • The training phase and the test phase use the same or identical asynchronous motor: this ensures that the asynchronous motor does not change after the training phase, if a change occurs, the anomaly detector must relearn.

[0024] So far, only MCSA solutions in combination with supervised learning are known. Additionally, "classical" MCSA uses FFT. The resulting disadvantages and problems are solved by the present invention:

[0025] - The disadvantage of supervised learning is that the training must use "good" and "bad" data. In customer equipment, it is usually not possible to record "bad" data, i.e. the fault state, for training, because for this the motor must be deliberately damaged. If the data recording takes place at the manufacturer, the manufacturer must provide data for a large number of different motors. Furthermore, the data recorded at the manufacturer do not take into account the customer's usage environment, which can lead to false detections. In contrast, the anomaly detector as in the present invention can be trained "unsupervised", i.e. using only "good" data.

[0026] - "Classical" MCSA, i.e. frequency analysis, uses discrete frequencies that arise from the physical properties of the motor. Faults such as misalignment change the physical properties of the motor ("eccentricity") only to a very low extent, so that such faults are difficult to identify. In contrast, the time-domain analysis as in the present invention can even reliably identify such faults.

[0027] - For a high-quality "classical" MCSA, i.e. frequency analysis, a high frequency resolution is required. This, however, requires a long recording time during which the system must remain in steady state. Furthermore, very high-dimensional feature vectors are generated, which impair the performance of the anomaly detector due to the "curse of dimensionality". These problems are avoided by the time-domain analysis as in the present invention.

[0028] Advantageous and improved embodiments of the present invention are given in the dependent claims.

[0029] According to a preferred embodiment of the method for generating the SOM, first in a first step a time series I(1...T) of T current values I t of the current I flowing through the asynchronous motor due to the supply voltage U applied to the asynchronous motor is acquired, during which the asynchronous motor is in a good state and the slip s of the asynchronous motor lies within a predetermined slip tolerance band i around the slip value s . Thus, the slip s lies within the value interval . Thereafter, in a second step a training vector , where the T vector elements are the T current values ​​I of the detected time series I(1...T) t Formed, and the slip value s i Assigned to training vector Under different slip values, the above two steps are repeated one or more times to obtain a set of training vectors as a training data set. Generate X times Y neurons N xy A matrix where each neuron N xy Assign weight vectors G of length T respectively xy . The training vectors from the training data set Assigned to neuron N xy , these training vectors are related to the neuron N xy The weight vector G xy Vector distance Do not exceed the predetermined distance limit . And adjust the neuron N xy The weight vector G xy , so that the value assigned to neuron N xy The weight vector G xy and assigned to the neuron N xy The training vector Vector distance are the smallest respectively.

[0030] According to a preferred design of the method for condition monitoring of an asynchronous motor, analysis and determination of anomalies are achieved by at least one trained AI function of the anomaly detector. Thus, the method utilizes a machine-based approach to detect anomalies. Current changes that occur in reality can be very diverse, as there are almost infinitely many different possibilities for the construction and elements of the power grid and drive system. Therefore, it can be difficult to fully define in advance all the criteria for when an anomaly exists, that is, to be able to clearly determine whether an anomaly exists for every possible current change. In this regard, the AI ​​function is advantageous because it can continuously learn and develop; therefore, the AI ​​function can apply the learned criteria to current changes that occur in reality and that the operator has not even thought of in advance, thereby allowing classification as to whether an anomaly exists.

[0031] By applying a trained function for artificial intelligence (AI function) to a time series of measured current values, function output data containing information about abnormalities in the asynchronous motor is generated. The trained function for abnormality identification can extract information about the quality of the asynchronous motor's operating state from the time series of measured current values.

[0032] A trained function is a function trained by a machine learning algorithm (ML = machine learning). Training is to be understood generally as an optimization of a mapping of input parameters of a parameterized system model, for example a neural network, to one or more target parameters. The mapping is optimized according to predefined criteria which have been learned and / or are to be learned in the training phase. For example, in a pattern recognition model, the percentage of successful pattern recognition can be used as a criterion. The training structure can include the network structure of the neurons of, for example, a neural network and / or the weights of the connections between the neurons which are structured by the training such that the criteria are fulfilled as well as possible.

[0033] In the present embodiment, a neural network (abbreviation: NN) is trained according to predetermined time current profiles in order to recognize motor abnormalities. In order to be able to distinguish a specific current profile pattern from other current profile patterns, the properties of the current profile pattern must be described and mathematically mapped. The higher the accuracy of the description of the pattern, the more information can be evaluated and the more reliable the pattern recognition run is.

[0034] In order to measure the motor quality, a current profile data set is formed and used for training the anomaly detector.

[0035] The trained function is part of the pattern recognition device. For configuration purposes, the pattern recognition device has one or more trainable functions, i.e. algorithms which can be executed in the computing module of the pattern recognition device, which implement a machine learning method in order to optimize the pattern recognition by training. These trainable functions can be trained using known machine learning standard methods in order to recognize the predetermined time current profiles in the input files as accurately as possible again.

[0036] Generally, the more training data, i.e. files available for training, the more successful the training is. In particular, the training data should representatively cover as large a range as possible of possible configurations of the asynchronous motor. This includes in particular training data on different slip values of the asynchronous motor, i.e. in different load states.

[0037] As described above, according to the application, a neural network NN is trained in order to recognize motor abnormalities. The neural network NN is trained using the SOM algorithm. Generally, a trained function imitates cognitive abilities which associate a person's understanding with that of other persons. In particular, a trained function is enabled to adapt to new situations and to find and deduce patterns by training based on training data.

[0038] Generally speaking, the parameters of a trained function can be adjusted through training (i.e., learning). This can typically be done using supervised learning, semi-supervised learning, unsupervised learning, reinforcement learning, and / or active learning. Furthermore, representation learning, also known as "feature learning," can be used. In particular, the parameters of a trained function can be iteratively adjusted through multiple training steps. In contrast, the present invention utilizes unsupervised learning.

[0039] According to a preferred embodiment of the present invention, characteristic data are determined from the measured current values, which are used to train the anomaly detector. Determining characteristic data from the measured current measurement sequence, such as the median value, minimum and / or maximum value, mean value, expected value of the distribution, and variance, is generally accompanied by a reduction in the amount of data, which significantly simplifies and accelerates data processing: the anomaly detector then only needs to process the characteristic data, rather than the complete measured current measurement sequence.

[0040] According to the preferred design of the present invention, select , where E max is the training vector used as the training data set All quantization errors of the group The maximum value of . Therefore, the "outlier bound" is constructed as the maximum deviation of the "good" data. The quantization error Ex exceeds the value E max The measurements were then classified as abnormal.

[0041] According to the preferred design of the present invention, select

[0042]

[0043] Among them, E max is the training vector used as the training data set All quantization errors of the group The maximum value, E min is the minimum value, L is determined as the sensitivity level , and among them, represents the value such that (100 * L)% of the training examples have anomaly scores less than this value. The following situations may occur:

[0044] Case 1: L = 0. This value results in a threshold value E Threshold = E max Therefore, the maximum error during training defines the threshold for anomaly detection (as described above).

[0045] Case 2: L > 0. The system should be "more sensitive". For L = 1, we get E Threshold = E mini.e. the system classifies each training vector of the training data set as abnormal. Thus, the system is maximally sensitive. In general, the sensitivity can be determined between 0% and 100%.

[0046] Case 3: L < 0. The system should be "less sensitive". For this, a value between -1 and 0 is determined so that the threshold E Threshold > E max The value L = -1 leads here to a threshold of 2*Emax - Emin. Obviously, in this case the histogram is mirrored at the point E = Emax.

[0047] Since the optimal threshold E Threshold is calculated and set from the desired level of sensitivity L, the system can be used for anomaly detection.

[0048] According to a preferred design of the application, the parameters N and p are defined such that if at least (100 * p)% of the last N time windows are classified as abnormal by the anomaly detector, an anomaly of the asynchronous motor is determined, otherwise the motor is assumed to be in a normal state.

[0049] According to a preferred design of the application, for the monitoring a pointer representation is used, in which the quantization error of the feature vector belonging to the current time window is shown with a pointer that can be moved along a scale. This pointer is dynamically moved with each occurring time window. An additional limit value is defined from which an increasing abnormal value already exists, but which has not yet been classified as faulty. Thus, a visualization of the abnormal size is achieved: the "green area" is defined by the bandwidth of the "good data"; higher abnormalities indicate motor problems, the greater they are, the more serious the problem is: yellow or red.

[0050] According to a preferred design of the pointer representation, the values ( ) and are chosen. Thus, an additional limit value is defined which indicates the error from which an increasing abnormal value already exists, but which has not yet been classified as faulty. In this way a warning range can be defined in which the user is already warned before the anomaly comes.

[0051] A further aspect of the application is a device for condition monitoring of a drive train having an asynchronous motor operated on an electrical network. The device has a sensor for detecting a current flowing through the asynchronous motor as a result of a voltage applied to the motor. The device has a first sensor unit for detecting a time series of current values of the current flowing through the asynchronous motor as a result of a supply voltage applied to the asynchronous motor. The device has a second sensor unit for determining a slip value of a slip of the asynchronous motor as it had during detection of the time series. The device has a computing unit which is adjusted to generate a feature vector from the current values of the detected time series if the determined slip value already existed when one of the training vectors was generated according to the method for generating a SOM; to assign to the feature vector a neuron Nx of the SOM and a weight vector Gx assigned to the neuron Nx, which weight vector was formed according to the method for generating a SOM; to calculate a quantization error Ex for the feature vector ; and to determine an anomaly of the asynchronous motor if the quantization error Ex exceeds a fixed predetermined threshold value E Threshold . The device further has an output unit for reporting the anomaly.

[0052] A further aspect of the application relates to a computer program comprising instructions for causing a device for condition monitoring of an asynchronous motor to perform a method for condition monitoring of an asynchronous motor when the program is executed on a computer.

[0053] A further aspect of the application is a computer program having instructions for causing a computer to perform a method for generating a SOM according to the application when the program is executed on the computer. BRIEF DESCRIPTION OF DRAWINGS

[0054] The application is explained below with reference to the accompanying drawings. In the drawings, which are schematic and not to scale,

[0055] Figure 1 a first design of a device for condition monitoring of a drive train according to the application is shown;

[0056] Figure 2 a second design of a device for condition monitoring of a drive train according to the application is shown;

[0057] Figure 3 a flow chart of a design of a method for generating a SOM for condition monitoring of an asynchronous motor according to the application is shown;

[0058] Figure 4 a flow chart of a design of a method for condition monitoring of an asynchronous motor according to the application is shown;​

[0059] Figure 5 The design of SOM is shown;

[0060] Figure 6 shows a histogram of quantization errors for normal and fault states for a first case;

[0061] Figure 7 shows the confusion matrix of the anomaly detector;

[0062] Figure 8 shows a histogram of quantization errors for normal and fault states for a second case;

[0063] Figure 9 A pointer meter for dynamically monitoring anomaly identification is shown. DETAILED DESCRIPTION

[0064] Figure 1 The device is shown with an asynchronous motor M, which is electrically connected via a supply line 11 to a voltage source 10, which supplies a voltage with a frequency f supply Supply voltage U supply The current I flowing through the power supply line 11 is measured by a current sensor 14, such as a shunt, a current transformer or a Hall effect sensor. The current value is respectively measured at a scanning frequency f A [1 / T] Time T A [T] records. Therefore, the number of sample values ​​is N A = f A *T A For example, when f A = 3200 Hz and T A =1.28 seconds, we get N A = 4096 = 2 12 The N sample values ​​recorded by the current sensor 14 A The current measurement values ​​are sent from the current sensor 14 to the calculation unit 16. A The current values ​​are respectively formed by the calculation unit 16 into a value of length LF = N A The eigenvector of .

[0065] The method steps for determining the slip s of the asynchronous motor M are also 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 determining the slip s from the current measurement values. This software is executed by the processor 18.

[0066] Input values, such as the number of neurons, can be transmitted to the computing unit 16 via an input / output unit 22 (eg, a PC) connected to the computing unit. Upon detecting an anomaly, the computing unit 16 sends a corresponding message to the input / output unit 22.

[0067] Figure 2 An alternative arrangement is shown, in which an asynchronous motor M is designed as the driving machine of a drive train D, which also includes a mechanical transmission G and a working machine W. The torque provided by the asynchronous motor M is transmitted to the working machine W via the transmission G. The working machine W can be, for example, a conveyor belt, rollers, or drums. In addition to the asynchronous motor M of the drive train D, other electrical loads 24 are also connected to the power grid N.

[0068] Figure 3 A flow chart shows one embodiment of the method according to the invention for generating an SOM for condition monitoring of an asynchronous motor M, to which a supply voltage U is applied.

[0069] In a first step 310 , T current values ​​I of the current I flowing through the motor M are detected. t , which current flows through the asynchronous motor M due to the supply voltage U applied to the asynchronous motor M. Thus, a time series I(1...T) of T current values ​​It is obtained, during which the asynchronous motor M is in a good state. During the recording phase, the current value I describing the "good state" of the motor is recorded t These current values ​​I t This is then used to train anomaly detectors. The following aspects are crucial:

[0070] - The system should record the slip value range [s min , ..., s N ] as many load states or slip values ​​s as possible i For each record, the corresponding slip value s i Should be used as a "label" L s carry.

[0071] - If a load state is recorded, the recording duration T for the load state (= slip) A Approximately constant: L s = constant. It is meaningless to record the current value in the case of large load changes (=slip changes), since otherwise a superposition of different states and thus different spectra would occur.

[0072] - The recording phase ends when a sufficiently large change in the slip value has been recorded or when the maximum time specified for the recording has been reached.

[0073] Step 320: From the detected time series current value I tgenerating training vectors , the slip value s i of the slip s of the asynchronous motor .

[0074] Step 330: repeating steps 310 and 320 for different slip values s i to obtain a set of training vectors as training data set.

[0075] Step 340: generating a matrix of X times Y neurons, wherein each neuron is assigned a weight vector G of length T.

[0076] Step 350: assigning training vectors of the training data set to the neurons, which training vectors have a vector distance, more precisely a vector distance to the weight vector G of the neurons, of not more than a predetermined distance limit.

[0077] Step 360: adjusting the weight vectors G of the neurons such that the distance of the weight vectors G to the training vectors assigned to the neurons is minimized.

[0078] The recording phase is followed by a learning phase. During this learning phase, the recorded and with the slip label L s labeled training vectors are learned into the SOM.

[0079] Figure 4 A flow chart showing a design scheme of a method for condition monitoring of an asynchronous motor M on which a supply voltage U is applied, using a SOM according to the invention is shown.

[0080] Step 410: detecting a time series I(1...T) of T current values of a current I flowing through the asynchronous motor M due to a supply voltage U applied to the asynchronous motor M.

[0081] Step 420: determining a slip value s i of a slip s of the asynchronous motor M which the asynchronous motor M had during detection of said time series I(1...T).

[0082] Step 430: if the determined slip value s i lies within a predetermined slip tolerance band , one of the training vectors generated according to the method for generating a SOM already exists, generating a feature vector Fx from the current values of the detected time series I(1...T).

[0083] Step 440: Assign the neuron Nx of the SOM and the weight vector Gx assigned to the neuron Nx to the feature vector Fx.

[0084] Step 450: Calculate the quantization error Ex for the feature vector Fx, .

[0085] Step 460: If the quantization error Ex exceeds a fixed predetermined threshold E Threshold : , then determine that the asynchronous motor M is abnormal.

[0086] Step 470: In the case of determining abnormality, report the abnormality.

[0087] Figure 5 Four different SOMs are shown. The SOM is a two-dimensional visualization stretched by the x-axis and the y-axis.

[0088] The idea of the SOM is to generate a grid of X * Y neurons N. Each neuron N contains here a weight vector, which has the same length as the feature vectors. In this case, the number of neurons (X * Y) is determined as a compromise between data volume / computational effort and the desired reliability of the anomaly detector. In tests that have been carried out, 20 x 20 = 400 neurons have produced very good results.

[0089] The parameter X is plotted on the x-axis and the parameter Y on the y-axis, wherein the size of the SOM map is defined by X and Y. The product X * Y gives the number of neurons represented as cells in the SOM. The left column labeled "X = 20" is two SOM maps shown by X = Y = 0,..., 19. The right column labeled "X = 25" is two SOM maps shown by X = Y = 0,..., 24. The left column has N = 20 * 20 = 400 neuron cells and the right column has N = 25 * 25 = 625 neuron cells. Both columns are divided horizontally into two rows: the upper row with "AR = 64" shows the SOM maps generated using an autoregressive AR with a lag Lag = 64, i.e. 64 feature vector elements per feature vector; the lower row with "AR = 512" shows the SOM maps generated using an autoregressive AR with a lag Lag = 512, i.e. 512 feature vector elements per feature vector. The SOM maps show either unfilled neuron cells or neuron cells filled with "good" or "faulty"; it is also possible here for a "double filling" of neuron cells to occur. The darker the color of a neuron cell, the greater the "distance" to the neighboring neuron cells. The optimum case is that as many neuron cells as possible are filled and as little "double filling" as possible.

[0090] Learning an SOM is an iterative process. Initially, the weight vectors contain random values. The learning process is then performed in such a way that similar training vectors from the training data set are assigned to the same neuron. In this context, "similar" should be understood in the sense of vector distances, i.e., depending on the application, for example, Euclidean distance, cosine distance, or Manhattan distance can be used.

[0091] Since the generated feature vector Number Usually much larger than the number of neurons, so one neuron represents multiple feature vectors. During training, the weight vector G of each neuron is adjusted so that the weight vector G is consistent with the training vector assigned to the neuron. The distance is minimized. In the optimally trained SOM, each neuron is assigned a training vector. The trained SOM then forms a compact representation of the training data distribution.

[0092] In order to be able to use the trained SOM as an anomaly detector, a function is required, which is the feature vector Assign the degree of abnormality to this eigenvector. To this end, for each eigenvector Calculating "quantization error" :

[0093]

[0094] For all ν, that is, all eigenvectors. Here, is the vector distance defined previously, is the corresponding weight vector of the "optimal" neuron, i.e., the neuron whose weight vector G has the smallest distance from the eigenvector. The quantization error E describes how well the corresponding eigenvector can be mapped to the data distribution used to train the SOM. A high error indicates that the eigenvector is very different from the training data and therefore describes unusual motor behavior.

[0095] Then determine the minimum error over all feature vectors in the training data set

[0096]

[0097] and maximum error

[0098] .

[0099] In the test phase following the learning phase, in which the device operates as an anomaly detector, the following steps are performed similarly to those in the recording phase:

[0100] - Again for time T A Record the sampled values.

[0101] - Check whether the load state of the motor during this time is "steady-state," i.e., without significant load changes, and whether slip is present during the recording in the training phase. Suitable tolerances for slip can be defined here, since slip determination cannot be error-free, or during the training phase, it can be checked which "slip errors" could lead to "anomalies."

[0102] - After a successful check, calculate the eigenvector F from the scanned values x .

[0103] - For this feature vector, SOM provides the associated neuron and the weight vector G corresponding to the neuron x .

[0104] - Calculate the quantization error from these values .

[0105] The simplest variant of the anomaly detector can be implemented as follows: define a fixed threshold E for the quantization error Threshold If the threshold is exceeded, there is a high probability that an anomaly exists. Obviously, choose E Threshold = E max , 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 becomes too large, there is an anomaly.

[0106] However, this simple variant has drawbacks, which are explained by experimental results using real motor data. To this end, a fault-free system (motor) ("good") is trained with 100 feature vectors. This training is performed for different load conditions. In a test phase, these 100 feature vectors are tested in the "good" state. The system is then deliberately "damaged," for example, with a motor misalignment ("anomaly"). Testing is performed only under the previously trained load conditions. A total of 200 additional feature vectors are tested.

[0107] Figure 6 The quantized error of the feature vector is shown, here with "error" on the x-axis. The number (frequency) of candidates (= scenarios) with the corresponding error is plotted on the y-axis. There are a total of 100 "good scenarios" (NGD=100) and 200 bad scenarios (NAD=200); therefore, the sum of all bars equals 100+200=300. In this case:

[0108] -E ν,Train Denotes the error of the training vectors generated during the training phase. The data is denoted as “good” and unshaded. It can be clearly seen that the errors follow an unknown distribution, but the maximum error is approximately 0.16, which also corresponds to the detection threshold.

[0109] -Eν,Test The error in the feature vectors generated during the test phase is shown. The data is shaded as "anomalies." It's also clear here that the errors follow an unknown distribution, but all errors are > 0.6, so anomalies are clearly identified for each state. The histogram shows the number of candidates / scenarios that fall within the corresponding anomaly.

[0110] Figure 7 The corresponding good truth matrix, also called the confusion matrix, is shown for the anomaly detector. All 100 training states with "good" motors were identified as "good" (or "healthy"), and the 200 test states were also correctly assigned.

[0111] Figure 8 Shows something like Figure 6 The histogram is used for applications where faults only cause very small changes in the current signal. Here, 900 "good" vectors are used for training and testing, and another 1,800 "abnormal" vectors are used for testing. Since the "good" and "abnormal" histograms cannot be completely separated, two possible types of misclassifications are obtained by the anomaly detector:

[0112] - False positive: Anomaly detector identifies a normal state as a fault.

[0113] - False negative: The anomaly detector identifies a faulty state as normal.

[0114] The importance of these two error types can vary depending on the application. In certain applications, it's important to generate as few false alarms (false positives) as possible, because, for example, each alarm results in a manual machine inspection, which is costly. In other applications, it's more important to indicate the problem as early as possible so that a shutdown can be scheduled as far in advance as possible.

[0115] In order to ensure the reliability of the anomaly detector, a fixed threshold for identifying anomalies is first determined. The appropriate threshold should be located at E max In this case, a smaller value results in a higher sensitivity of the anomaly detector, i.e., in principle, more alarms are generated, and thus more false alarms may be generated. A higher threshold value results in fewer alarms, thus potentially reducing false alarms, but also in a potentially lower recognition rate. The error during the learning phase is and There is a fixed distribution between , which is visualized by a histogram. To determine the threshold, first the user of the system defines a “sensitivity level” The threshold is calculated as

[0116]

[0117] where The return value for which it is true that for (100 * L)% of the training examples the anomaly score is less than this value. The following can occur here:

[0118] Case 1 : L = 0. This value leads to the threshold E Threshold = E max . Thus, the maximum error during training defines the threshold for anomaly detection (as described above).

[0119] Case 2 : L > 0. The system should be "more sensitive". For L = 1, one obtains E Threshold = E min , i.e. the system would classify every feature vector in the training data set as an anomaly. Thus, the system is maximally sensitive. In general, the sensitivity can be determined between 0% and 100%.

[0120] Case 3 : L < 0. The system should be "less sensitive". For this, a value between -1 and 0 is determined so that the threshold can be reached. The value L = -1 leads here to the threshold = E = E

[0121] Since the optimal threshold E Threshold is calculated and set from the desired sensitivity level L, this system can be used for anomaly detection. In order to additionally avoid false alarm rates, alarms are suppressed for isolated occurring anomalies. For this, two further parameters N and p are defined. If at least (100 * p)% of the last N time windows are classified as anomalies by the anomaly detector, an alarm is output. Otherwise the normal state of the electric motor is assumed. For example, with a recording time T A = 1.28 seconds and the values N = 10 and p = 50, an alarm is only output when at least half of all time windows in the last 12.8 seconds are classified as anomalies.

[0122] Figure 9 A "pointer instrument" is shown which can be used for monitoring the system. The "pointer instrument" can be used on a monitor, network or cell phone.

[0123] On the axis 81, the quantized error of the feature vector belonging to the current time window is shown with the pointer 80. The pointer 80 moves dynamically with each occurring time window. Thus, an additional limit value is defined which indicates from which error an increased anomaly value has already existed, but which has not yet been classified as faulty. For example, one can choose and The left area 82 ("green area") corresponds to the error region of the training data. Max The pointer 80 moves to the middle area 83 (“warning area”) as soon as the threshold E is exceeded. Threshold , the pointer 80 moves to the right area 84 ("red area"). Threshold If it is exceeded frequently enough (see above), an alarm is additionally output.

Claims

1. 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, in, The current value I of the current I is t A time sequence I(1...T) of currents flowing through the asynchronous motor (M) due to the supply voltage U applied to the asynchronous motor (M), during which the asynchronous motor (M) is in a good state, forms a plurality of training vectors for generating an SOM.

2. The method according to claim 1, comprising the steps of: a) Obtain T current values ​​I of current I t A time series I(1...T) of the current flowing through the asynchronous motor (M) due to the supply voltage U applied to the asynchronous motor (M), during which the asynchronous motor (M) is in a good state and the slip (s) of the asynchronous motor (M) is at the slip value s i Predetermined slip tolerance band around Inside; b) Forming a training vector Fν with T vector elements, where The T vector elements are composed of the T current values ​​I detected in the time series I(1...T) t Formed, and the slip value s i Assigned to training vector ; c) At different slip values ​​(s i1 , s i2 , s i3 , ...), repeat steps a) and b) one or more times to obtain a set of training vectors As a training data set; d) Generate X times Y neurons N xy A matrix where each neuron N xy Assign weight vectors G of length T respectively xy ; e) The training data set The training vector Assigned to neuron N xy , the training vector and neuron N xy The weight vector G xy Vector distance Do not exceed the predetermined distance limit ; f) Adjust neuron N xy The weight vector G xy , so that the value assigned to neuron N xy The weight vector G xy and assigned to the neuron N xy The training vector Vector distance are the smallest respectively.

3. A method for condition monitoring of an asynchronous motor (M) to which a supply voltage (U) is applied, using a SOM generated by the method according to any of the preceding claims, comprising the following steps: a) detecting a time series I(1 . . . T) of T current values ​​of a current I which flows through the asynchronous motor (M) as a result of a supply voltage (U) applied to the asynchronous motor (M); b) determining a slip value s of the asynchronous motor (M) during the detection of the time sequence I(1 . . . T) i ; c) If the determined slip value s i By predetermined tolerance When generating SOM training vectors If one of them already exists, a feature vector Fx is generated from the current values ​​of the detected time series I(1...T); d) The neurons N of the SOM xy and assigned to the neuron N xy The weight vector is assigned to the feature vector Fx; e) calculating the quantization error Ex for the eigenvector Fx, ; f) If the quantization error Ex exceeds a fixed predetermined threshold E Threshold : , it is determined that the asynchronous motor (M) is abnormal; e) Report anomalies when identified.

4. The method according to claim 3, wherein The analysis and determination of anomalies are performed by an anomaly detector (16) having at least one trained AI function (18).

5. The method according to claims 3 and 4, wherein: choose , where E max is the training vector used as the training data set All quantization errors of the group The maximum value of .

6. The method according to any one of claims 3 to 5, wherein the ， Among them E max is the training vector used as the training data set All quantization errors of the group The maximum value, E min is the minimum value, L is determined as the sensitivity level , And among them, The following value is represented, that is, it holds: for (100 * L)% of the training vectors, the anomaly score is less than this value.

7. The method according to any one of claims 3 to 6, in, The parameters N and p are defined such that an abnormality of the asynchronous motor is determined if at least 100*p% of the last N time windows are classified as abnormal by the anomaly detector, otherwise the motor is assumed to be in a normal state.

8. The method according to any one of claims 3 to 7, wherein For monitoring, a pointer representation is used, wherein the quantization error of the characteristic vector belonging to the current time window is shown on the pointer scale with a pointer, wherein the pointer is moved dynamically with each occurring time window, and wherein additional limit values ​​are defined , the additional limit value indicates, starting from which error an increasing outlier value is already present but which has not yet been classified as a fault.

9. The method of claim 8, wherein L>0 is selected and .

10. A device for monitoring the condition of an asynchronous motor (M) using a SOM generated by the method according to claim 1 or 2, to which a supply voltage (U) is applied, comprising: The first sensor unit is configured to detect a time series (I) of T current values ​​of the current (I) t ), the current flowing through the asynchronous motor (M) due to the supply voltage (U) applied to the asynchronous motor (M); The second sensor unit is used to determine the asynchronous motor (M) when detecting the time sequence (I t ) during the period of time, the slip value of the asynchronous motor (M); A calculation unit adapted to: - If the determined slip value passes the predetermined tolerance in generating the training vector If one of them already exists, then from the detected time series (I t ) generates a characteristic vector (Fx) of the current value, - assigning the neuron Nx of the SOM and the weight vector Gx assigned to the neuron Nx to the feature vector (Fx), -Calculate the eigenvector The quantization error Ex: , - If the quantization error Ex exceeds a fixed predetermined threshold E Threshold , it is determined that the asynchronous motor (M) is abnormal; and - Output unit for reporting exceptions.

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

Images