Method for monitoring state of asynchronous motor

Abstract

The invention relates to a method for monitoring the state of an asynchronous motor (AM) having a slip ratio (s) and to which a supply voltage (U) having a supply frequency (f) is applied, the method comprises the following steps: defining two or more different operating ranges (B) of the asynchronous motor (AM) as a function of the slip ratio (s) and of the supply frequency (f); each defined operating range (B) is associated with at least one associated anomaly detector (Ai-1, Ai, Ai + 1), which is suitable for detecting a fault state (E) of the asynchronous motor (AM) in the respective operating range (B); -ascertaining a current operating range (B) of the asynchronous motor (AM); activating at least one anomaly detector (A) associated with the current operating range (B) and monitoring the state of the asynchronous motor (AM) by means of the activated anomaly detector (A); and-generating a corresponding signal in the event that the activated anomaly detector (A) detects an anomaly.

Description

[0001] The invention relates to a method for condition monitoring of an asynchronous motor.

[0002] Asynchronous motors are very common in industry due to their cost-effectiveness and robust design. They are used in a wide range of applications, such as fans, saws, and pumps, ranging from relatively small, multi-watt devices to megawatt-class equipment. Consequently, asynchronous motors are one of the largest consumers of electrical energy generated worldwide. The widespread use of asynchronous motors offers enormous potential for increased availability and reliability, as well as the associated savings in time and money. Terms such as "condition monitoring" (CM) and "predictive maintenance" (PM) are often used in conjunction with this.

[0003] Asynchronous motors (ASMs) operate according to the principle of electromagnetic induction, where the interaction between currents or magnetic fields in the stator and rotor generates torque that drives the rotor shaft. Motor faults such as eccentricity, particularly air gap eccentricity (when the stator and rotor axes of rotation do not completely coincide), broken rotor bars, bearing damage, or misalignment, can alter the amplitude of characteristic fault frequencies at constant slip. The fault frequencies themselves depend on the slip; changes in slip alter the frequency profile of the fault frequencies. Therefore, motor faults can be identified by changes in the amplitude of these characteristic frequencies in the frequency spectrum.

[0004] To this end, so-called Motor Current Signature Analysis (MCSA) can be used, see, for example, EP 3961230 A1 (Siemens; Zettner, Jürgen) March 2, 2022. For example, Fast Fourier Transformation (FFT) is used to detect and quantify motor faults and operating states. Classic MCSA is used in quasi-static situations, i.e., at nominally constant speed. Extensions to MCSA include analysis using short-time FFTs or wavelets for time-resolved process analysis. However, machine learning (or artificial intelligence) methods also exist.

[0005] Anomaly detection in the prior art relies on a learning phase on the asynchronous motor to be monitored, during which the asynchronous motor is in a so-called "good state," i.e., a state in which the asynchronous motor operates "normally," i.e., without major faults (what is considered "normal" and what is considered a major fault must be defined in advance by the operator of the asynchronous motor); anomaly detection in the detection phase is based on the features that the anomaly detector has been trained on as "normal" during its learning phase. Typically, during the learning phase (also called the training phase), the good state is learned under different operating conditions, with the expectation that the mathematical model of the anomaly detector covers all operating states of the good state in the feature space.

[0006] Variable frequency drives and drive trains with mains motors experience variations in rotor speed and torque due to the driven load (= torque), temperature changes, aging, or control / regulation. As described in the MCSA theory, the frequency spectrum of an asynchronous motor (the position of the frequency line) depends significantly on the load state, i.e., the slip s. Furthermore, as can be seen from the classical MCSA, the frequency line also varies with the frequency of the supply voltage applied to the asynchronous motor ("mains frequency f supply ”). MCSA primarily distinguishes between motor-specific spectral lines that depend on slip, such as main slot harmonics (PSH), various slip-dependent fault lines, and motor-specific winding harmonics (WH), which are multiples of the mains frequency.

[0007] For PSH, the following applies:

[0008] f PSH = [k•(R / p)•(1−s)±ν] • f supply Formula (1)

[0009] For typical faults:

[0010] i) Broken Bars (BB):

[0011] f BB1 = (1 ± 2k•s) • f supply Formula (2)

[0012] ii) Static / dynamic rotor eccentricity (see ISO 20958:2013)

[0013] f ecc1 = 1±(k / p)•(1−s) • f supply Formula (3)

[0014] f ecc2 = [(R±n d )•(k / p)•(1−s)±ν] • f supplyFormula (4)

[0015] iii) Bearing damage of rolling bearings

[0016] f o = [(N b / 2)•(1−(D b / D p )•cos(β)] • f rot Formula (5)

[0017] f i = [(N b / 2)•(1+(D b / D p )•cos(β) / D p ) • f rot Formula (6)

[0018] in,

[0019] f BB1 is the frequency when the rotor bar breaks

[0020] f supply is the frequency of the power supply voltage, for example 50 Hz

[0021] f ecc1 is the eccentric frequency (family 1)

[0022] f ecc2 is the eccentric frequency (family 2)

[0023] f PSH Main slot harmonic frequency (= PSH)

[0024] f o is the frequency of the bearing outer ring (o = outer)

[0025] f i is the frequency of the bearing inner ring (i = inner)

[0026] f rot is the rotation frequency of the rolling bearing

[0027] k is a natural number

[0028] p is the number of pole pairs

[0029] R is the number of rotor bars

[0030] s is the slip rate

[0031] ν is an odd number 1, 3, 5, 7, ... = harmonic order of the power frequency

[0032] n dis an index depending on the type of eccentricity (static: n d = 0; dynamic: n d = 1, 2, 3, 4, ...)

[0033] N b is the number of rolling elements

[0034] D b / D p is the characteristic diameter of the rolling bearing

[0035] ß is the contact angle

[0036] Depending on the number of pole pairs P and the number of rotor bars, different modulation artifacts appear in the frequency spectrum, depending on the slip s of the asynchronous motor. A change in the load and, therefore, in the slip s, results in a superposition of different states and, consequently, different frequency spectra. Therefore, if the anomaly detector interprets different operating states (due to the varying power frequency f) during the learning phase, supply If a motor with a high torque and / or varying torque (asynchronous motor with slip) is used as a good state learning, then several motor-specific spectral lines and several frequency bands change with different operating states. Therefore, in the event of a fault during the detection phase, the probability of clearly finding the fault in the drive train is relatively low. This is because in some operating states, the amplitude of the motor-specific frequency and the amplitude of the fault-specific frequency are superimposed, see Figure 6 .

[0037] Figure 6 shows the horizontal time axis t, Figure 6 The lower area plots the motor speed v of the motor motor It can be seen that the motor under consideration is run twice in 12 different operating states 62 identified by different motor speeds, specifically starting from the highest speed to the lowest speed. The first run in these operating states is carried out in a good state 63, and the second run in these operating states is carried out in a faulty state 64. Figure 6 In the upper area of ​​, the frequency spectrum of the motor obtained by MCSA is also plotted on the time axis t. It can be seen that in some operating states, the amplitude 60 of the motor-specific frequency overlaps with the amplitude 61 of the fault-specific frequency.

[0038] This means that if all the behaviors of the good state are learned into the anomaly detector, then Figure 6 The presence of the 12 operating states shown and their corresponding amplitudes at the motor-specific frequencies is classified as "OK." If a fault is present that manifests itself at one of the fault-specific frequencies, separation from the OK condition is no longer necessary, since the amplitude of the fault frequency was once classified as the OK line, and in particular, the amplitude at the fault frequency may be smaller than the amplitude at the motor-specific amplitude.

[0039] Therefore, a method is needed by which a fault in a transmission chain can be clearly identified with a higher probability. Therefore, the technical problem to be solved by the present invention is to improve MCSA.

[0040] According to the invention, this object is achieved by a method having the features of claim 1 .

[0041] The method is used for condition monitoring of asynchronous motors. Due to their design, asynchronous motors exhibit slip during operation: slip is defined as the speed difference between the stator's rotating magnetic field (stator) and the rotor, typically expressed as a percentage of the stator's rotating magnetic field's speed. A mains voltage is applied to the asynchronous motor as its operating voltage. This mains voltage generates a current in the motor, which generates a magnetic field within the motor that causes the rotor to rotate. As alternating current, the mains voltage has a frequency, the so-called mains frequency.

[0042] According to the method, two or more different operating ranges of the asynchronous motor are defined as a function of the slip and as a function of the power supply frequency: the two different operating ranges thus differ in terms of the slip and the power supply frequency. Each defined operating range is assigned at least one associated anomaly detector, which is suitable for detecting a fault state of the asynchronous motor within the corresponding operating range. After determining which defined operating range the asynchronous motor is currently in ("current operating range"), at least one anomaly detector correspondingly assigned to the current operating range is activated. The state of the asynchronous motor is monitored by the activated anomaly detector. If the activated anomaly detector detects an anomaly in the asynchronous motor, a signal indicating the detected anomaly is generated.

[0043] The present invention proposes to use an anomaly detector specially trained for each operating range of the asynchronous motor in each operating range. The measures according to the present invention have the following advantages:

[0044] The present invention improves the accuracy of anomaly detection for each speed, torque, or operating point because the variance of features within the operating range is smaller. The smaller the "operating range," the lower the rate at which true faults are detected as "good," and the higher the accuracy: a smaller operating range can be better analyzed than a larger one.

[0045] - The invention avoids the situation where an anomaly detector responsible for all operating ranges recognizes all of them but “ignores” the faults that occur. This means that the probability of the second type of fault (false negative: “good report” despite the presence of a fault) is minimized.

[0046] - The number of features that need to be calculated at any operating point is minimal. The input values ​​of the anomaly detector, that is, the so-called features required to train the AI ​​for anomaly detection, can be obtained in various ways. The first method for obtaining input values ​​is a method for diagnosing the state of an electric motor to which an AC power supply voltage having a power supply frequency is applied, and the method has the following steps:

[0047] a) detecting a time series of current values ​​of a machine current flowing through the motor due to a power supply voltage applied to the motor;

[0048] b) Describing time series with a p-order autoregressive model ;

[0049] c) Form a feature vector with p vector elements, where the p vector elements are the p parameters a of the autoregressive model i formed; and

[0050] d) Using the generated feature vector as input value in the ML algorithm, that is, as training data or test data.

[0051] Another method of acquiring input values ​​is a method for diagnosing a state of an asynchronous motor to which a three-phase voltage having a power supply frequency is applied, the method comprising the following steps:

[0052] a) Time series of detection voltage U t and the current time series I of the three-phase motor current flowing through the asynchronous motor due to the voltage applied to the asynchronous motor t ;

[0053] b) Using FFT, based on the detected time series I t 、U t , calculate the current spectrum I(f,t) and voltage spectrum U(f,t);

[0054] c) Calculate one or more of the following time-varying spectral quantities 1-6:

[0055] 1) Current component I a (f,t),I b (f,t), where I a (f,t) and I b (f, t) is transformed by Park from I R,S,T to obtain;

[0056] 2) Total current amplitude A I (f,t) = √((I a (f,t)) 2 +(I b(f,t)) 2 )

[0057] 3) Phase angle φ between current components ab (f,t):φ ab (f,t) = arctan(I a / I b )

[0058] 4) Phase angle φ between current component and voltage component UI (f,t) = arctan(U(f,t) / I(f,t))

[0059] 5) Instantaneous spectral power P(f,t) = U(f,t)*I(f,t)

[0060] 6) Admittance X(f,t) = I(f,t) / U(f,t)

[0061] d) forming a feature vector from one or more time-varying spectral quantities; and

[0062] e) Using the generated feature vector as input to the ML algorithm, i.e., as training data or test data for the anomaly detector.

[0063] In a machine learning-based method for diagnosing the condition of an electric motor, input values, namely, so-called features required for training the AI ​​for anomaly detection, are obtained from current or voltage sampling values. For example, one of the two methods described above can be applied. In connection with this machine learning-based method for diagnosing the condition of an electric motor, it is recognized that features should be selected so as to contain as much information as possible. Even with a self-organizing map (SOM) using autoregressive features or FFT frequency features, if an arbitrary number of features or multiple amplitudes are used for training in noise equivalence, noise and artifacts, rather than fault characteristics, may be detected during training.

[0064] The size of each anomaly model A(q,r), such as the map size of a SOM, is reduced by the present invention because less training data exists for each restricted operating range. Consequently, training becomes easier and predictions faster, and storage of the models for each operating range is more resource-efficient. Each individual model (anomaly detector) is more specific and smaller.

[0065] The following describes the slip rate s, rotor speed n, motor torque M and power frequency f S The relationship between them.

[0066] Rotating magnetic field speed n at the stator S (unit: revolutions per second) is the power frequency fS (unit: Hz) and the number of pole pairs p:

[0067] n S = f S / p(I)

[0068] Power frequency f S , also known as stator frequency or grid frequency, can also be simply represented by the letter f without the subscript S.

[0069] The slip s is the speed of the rotating magnetic field at the stator n S The difference between the rotor speed n(rotor) (in revolutions per second) and the rotor speed n(rotor), usually expressed as the rotating magnetic field speed n S The percentage is expressed as:

[0070] s = (n S - n) / n S <=> n = n S (1 - s)(II)

[0071] The greater the torque M required of the rotor, the greater the degree to which the rotor speed n is reduced, and the greater the slip s.

[0072] The torque M of the motor is a function of the slip rate s. The torque M of the asynchronous motor is related to the slip rate value s and the critical slip rate s. K and the step-out torque M K The relationship between is approximately given by the Kloß formula:

[0073] M = 2⋅M K / (s K / s + s / s K )(III)

[0074] In addition, the following relationships apply:

[0075] P el = η P mech (IV)

[0076] P el = 3 UI cos φ(V)

[0077] P mech = ω M = 2 π n M(VI)

[0078] in,

[0079] P el The active power of the input motor

[0080] η is the efficiency of the motor

[0081] Pmech The mechanical power output on the motor shaft

[0082] U is the power supply voltage

[0083] I is the motor current

[0084] φ is the phase shift angle

[0085] ω is the angular frequency

[0086] From formula VI, by substituting into formulas II, I, IV, and V in sequence, the motor torque M is obtained as:

[0087]

[0088] The present invention uses an anomaly detector for the voltage signal. This can be a SOM (self-organizing map) or an autoencoder that detects deviations from an undisturbed ("good") sinusoidal voltage waveform. To measure voltage quality / grid quality, a voltage feature dataset, such as a voltage spectrum or autoregressive features, is generated and used to train the voltage anomaly detector.

[0089] The concept behind a SOM is to generate a grid of N neurons (X x Y). Each neuron N contains a weight vector of length LF, which is equal to the length of the feature vector. Determining the number of neurons (X x Y) is a compromise between the amount of data / computational effort and the desired reliability of the anomaly detector. In tests conducted, 20 x 20 = 400 neurons produced very good results.

[0090] Like most artificial neural networks, self-organizing maps operate in two modes: training and mapping. In training mode, an input dataset (the "input space") is used to generate a map (the "mapping space") that is a low-dimensional representation of the input data. In mapping mode, the generated map is used to classify the measurement data, for example, as normal or abnormal.

[0091] Learning, or training, the SOM is an iterative process. Initially, the weight vectors contain random values. The learning process proceeds by assigning similar feature vectors in the training dataset to the same neurons. "Similar" is understood in this context in terms of vector distances; depending on the application, for example, Euclidean distance, cosine distance, or Manhattan distance can be used.

[0092] The goal of training is to represent a p-dimensional "input space" with p variables into a two-dimensional "mapping space." This "mapping space" consists of components called nodes or neurons, which are arranged on a hexagonal or rectangular grid with two dimensions. The number of nodes and their location on the grid are predetermined based on the objectives achieved through analysis and study of the "input space."

[0093] Each node in the "mapped space" is assigned a weight vector that defines the node's position in the "input space." The nodes in the "mapped space" remain anchored at their positions, while the weight vector approximates the input data. This means that the distance between the weight vector and the input data, such as the Euclidean distance, is minimized without destroying the topology generated by the "mapped space."

[0094] During the training phase, “similar” feature vectors are grouped. This similarity is defined here as the Euclidean distance, i.e., the smaller the distance between feature vectors, the more similar the feature vectors are. For each “group” of similar feature vectors, one or more representatives are selected, and the maximum distance D of each group is max .

[0095] After the training phase, the SOM generated in this way can be used to classify other "input space" vectors, i.e., investigations in the "input space", here the time series of current values, by identifying the following nodes, whose weight vector is closest to the "input space" vector, i.e., has the smallest distance measure, such as the Euclidean distance.

[0096] A feature vector generated according to the invention having p vector elements, which are obtained from a time series measured at the motor by an AR method, is assigned to one of the neurons of the SOM and to the weight vector assigned to the neuron.

[0097] In order to be able to use the trained SOM as an anomaly detector, a function is needed that assigns to the eigenvector Fν a degree to which this eigenvector represents an anomaly. For this purpose, a "quantization error" Eν is calculated for each eigenvector Fν:

[0098] Eν = | Fν – Gν|

[0099] for all ν, i.e. all eigenvectors. Where |...| is a predefined vector distance and Gν is the weight vector of the "best" 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 onto the distribution of the data used to train the SOM. A high error indicates that the eigenvector is very different from the training data and therefore describes an abnormal behavior of the motor. In practice, if the quantization error Ex exceeds a fixed preset threshold E Threshold :Ex > E Threshold , it is determined that the motor is abnormal.

[0100] After training, anomaly detection is performed as follows: a feature vector is generated from the input values ​​as described above. For each of these vectors, the Euclidean distance to the learned "group" is determined. The minimum value is determined from all distances. If this value is greater than D max , there is an exception.

[0101] Advantageous embodiments and developments of the invention are specified in the dependent claims.

[0102] According to a preferred embodiment of the present invention, a slip value range of the slip value occurring during the operation of the asynchronous motor and a frequency value range of the power supply frequency occurring during the operation of the asynchronous motor are determined as a prerequisite for the operating point-specific training of the anomaly detector according to the present invention. The determined slip value range is divided into Q slip sub-ranges, and the determined frequency value range is divided into R frequency sub-ranges (Q, R ∈ ). Accordingly, Q×R operating ranges are defined, which are defined by one of the slip rate sub-ranges and one of the frequency sub-ranges. Therefore, according to the present invention, instead of training one anomaly detector, Q×R anomaly detectors A(q,r) are trained; the anomaly detector A(q,r) has a feature space of dimension K for this purpose, i.e. K represents the number of features of the anomaly detector A(q,r). Each anomaly detector is effective for the assigned operating range. In the detection phase, the current operating range of the asynchronous motor is determined, and then the anomaly detector corresponding to the current operating range is activated, and the state of the asynchronous motor is monitored by the activated anomaly detector. In the case where the activated anomaly detector detects an anomaly, a signal indicating the anomaly is generated by the activated anomaly detector.

[0103] According to a preferred embodiment of the present invention, during the detection phase, after determining the current operating range of the asynchronous motor, the anomaly detector corresponding to the current operating range is activated, as well as one or more anomaly detectors corresponding to operating ranges adjacent to the current operating range of the asynchronous motor. The state of the asynchronous motor is monitored by the activated anomaly detectors. This has the advantage that each anomaly detector knows the motor-specific frequencies learned for its "own" operating point. In a good state, only these frequencies occur, or the good state is best described by the K features of R frequencies according to the training. In a fault situation, the amplitude changes at the fault-specific frequencies differ significantly from the amplitudes of the motor-specific frequencies, which are fixed at the current frequency for that operating point.

[0104] Another advantage of N anomaly detectors is that even if Figure 6As shown, a majority decision can be made even at certain points where the fault-specific and motor-specific frequencies overlap, meaning that a fault cannot be readily identified for a specific operating point. Anomaly detectors at adjacent operating points (where the degree of overlap varies) can then conclude with greater reliability that an anomaly is present. Even if the precise actual operating point cannot be identified due to frequency overlap, the majority of the decisions at adjacent operating points can still lead to a conclusion that a fault is present. There may also be faults that affect only a specific range of operating points (cavitation is an example: assume cavitation only occurs when the speed n satisfies: 3000 rpm > n > 2400 rpm). Therefore, to detect with high probability that a fault has occurred at the actual operating point (e.g., within the range of 2600-2650 rpm, but where the anomaly detector has an overlap between the fault line and the motor line), the majority decision of the adjacent anomaly detectors for that operating point range is decisive.

[0105] According to a preferred embodiment of the present invention, a plurality of anomaly detectors are provided for each defined operating range, wherein the detection results of the corresponding anomaly detectors are stored as result vectors. A majority decision is applied to the result vectors to determine whether an anomaly exists.

[0106] According to a preferred embodiment of the present invention, determining the current operating range of the asynchronous motor comprises the following steps: first, determining the slip value and the power supply frequency value during a time series, and determining the current operating range of the asynchronous motor based on the determined slip value and power supply frequency value.

[0107] According to a preferred embodiment of the present invention, the time series is divided into subunits of shorter duration, wherein a slip metric is determined in each subunit. The anomaly detector is activated only if the difference in the slip metric does not exceed a predetermined difference value; otherwise, the anomaly detector is not activated.

[0108] Since determining the slip or power frequency also requires a certain time span T, the operating point may change during the detection time. By dividing the time series T into ν sub-units Tν with shorter duration, the metric L for the slip can be determined in each sub-segment. s、ν In this case, the measure for the slip can be determined in the frequency domain from the position of the main slot harmonics (PSH) or in the time domain from cosφ.

[0109] If the metric L of the sub-segment s、ν If the two are not close enough to each other, the operating state is likely to change during T and no anomaly detector is "triggered". Alternatively, each anomaly detector A(n, m) is detected by s、νHowever, anomalies are only output when enough operating point specific anomalies have accumulated, which is unlikely during a change.

[0110] The invention is explained below with reference to the accompanying drawings, which are each shown schematically and not to scale:

[0111] Figure 1 A design scheme of a device for monitoring the state of an asynchronous motor according to the present invention is shown;

[0112] Figure 2 A design scheme of an artificial neural network is shown;

[0113] Figure 3 A flow chart showing a design solution of the method according to the present invention is shown;

[0114] Figure 4 The definition of different operating ranges of asynchronous motors is shown;

[0115] Figure 5 A design scheme in which multiple anomaly detectors are responsible for one operating range is shown.

[0116] Figure 1 An asynchronous electric motor AM (hereinafter also referred to as "motor" or "electric machine") is shown, which is electrically connected to the power grid N via a branch line 12. This branch line can be, for example, a three-phase power line that connects the three-phase power grid N to the terminals of the electric motor AM. The electric motor AM is designed as the drive of a drive train D, which also includes a mechanical gearbox G and a working machine W. The torque provided by the electric motor AM is transmitted to the working machine W via the gearbox G. The working machine can be, for example, a conveyor belt, rollers, or rolls. In addition to the electric motor AM of the drive train D, other electrical loads 20 are connected to the power grid N.

[0117] For example, a first sensor S1 arranged on the branch line 12 detects the voltage value U of the supply voltage of the motor AM. m , such as the voltage between two phases of the branch line 12 or the voltage between one phase of the branch line 12 and the reference potential, and transmit the detected voltage value U via the transmission device 10, such as a data cable m The second sensor S2, such as a shunt, a current converter or a Hall sensor, arranged on the motor M detects the current value I of the current flowing through the motor AM. m , and transmits the current value I m The data are sent to a device 15 for monitoring the state of the asynchronous motor AM.

[0118] The current I flowing through the power supply line 11 is measured by a current sensor 14 at a measurement rate of, for example, 3200 measurements per second. The supply voltage U applied to the asynchronous motor AM is detected on the power supply line 11 by a voltage sensor 15. The current measurement values ​​detected by the current sensor 14 and the voltage measurement values ​​detected by the voltage sensor 15 are transmitted by the sensors 14 and 15 to a computing unit 16. The computing unit calculates the slip and FFT(I) from the detected current and voltage measurement values. The other method steps for determining the slip s are also performed by the computing unit 16. To this end, the computing unit 16 includes 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. To perform the method according to the present invention, the current and voltage must be measured, and the slip, i.e., the actual speed, must be determined in order to determine the operating point and the associated slip-dependent fault frequency.

[0119] The device 15 for monitoring the condition of an asynchronous motor AM comprises an anomaly detector 16, a user interface 17 and a data memory 19. The anomaly detector 16 comprises a trained artificial intelligence function 18. The device 15 receives the collected voltage value U m And the collected current value I m and analyzes it via the trained artificial intelligence function 18. The artificial intelligence function 18 analyzes the collected measured values ​​U m , I m and...

[0120] defining two or more different operating ranges (B) of an asynchronous motor (AM) as a function of slip (s) and as a function of supply frequency (f);

[0121] - For each defined operating range (B), at least one associated anomaly detector (A i-1 、A i 、A i+1 ), the detector is suitable for detecting a fault state (E) of an asynchronous motor (AM) in a corresponding operating range (B);

[0122] - Confirm the current operating range (B) of the asynchronous motor (AM);

[0123] - activating at least one abnormality detector (A) correspondingly assigned to the current operating range (B), and monitoring the state of the asynchronous motor (AM) by means of the activated abnormality detector (A); and

[0124] - In the event that an abnormality is detected by the activated abnormality detector (A), a corresponding signal is generated.

[0125] The device 15 reports a fault in the asynchronous motor AM when an abnormality is detected. The reporting is achieved through a user interface, such as a display 17.

[0126] Figure 2 An embodiment of an artificial neural network 100 is shown, as well as its use in a trained artificial intelligence function 18. Alternative names for "artificial neural network" are "neural network", "artificial neural network" or "neural network".

[0127] The artificial neural network 100 has nodes 120, ..., 132 and edges 140, 141, 142, wherein each edge 140, 141, 142 is a directed connection between a first node 120, ..., 132 and a second node 120, ..., 132. Although the first node 120, ..., 132 and the second node 120, ..., 132 are usually different nodes 120, ..., 132, it is also possible that the first node 120, ..., 132 and the second node 120, ..., 132 are the same. For example, in Figure 6 , edge 140 is a directed connection from node 120 to node 123, and edge 142 is a directed connection from node 130 to node 132. Edges 140, 141, 142 from a first node 120, ..., 132 to a second node 120, ..., 132 are also referred to as "ingoing edges" (English: incoming edge) with respect to the second node 120, ..., 132, and "outgoing edges" (English: outgoing edge) with respect to the first node 120, ..., 132: an "incoming edge" of a node is a directed edge with that node as its destination, and an "outgoing edge" of a node is a directed edge with that node as its origin.

[0128] In this embodiment, the nodes 120, ..., 132 of the artificial neural network 100 are arranged in layers 110, ..., 113, wherein these layers may have inherent regularities introduced by the edges 140, 141, 142 between the nodes 120, ..., 132. In particular, edges 140, 141, 142 may only exist between adjacent layers of nodes. In the illustrated embodiment, there is an input layer 110 having only nodes 120, 121, 122 without incoming edges, an output layer 113 having only nodes 131, 132 without outgoing edges, and hidden layers 111, 112 between the input layer 110 and the output layer 113. In general, the number of hidden layers 111, 112 can be selected arbitrarily. The number of nodes 120, 121, 122 in the input layer 110 generally corresponds to the number of input values ​​of the neural network, and the number of nodes 131, 132 in the output layer 113 generally corresponds to the number of output values ​​of the neural network.

[0129] In particular, each node 120, ..., 132 of the neural network 100 can be assigned a (real number) as a value. (n) i represents the value of the i-th node 120, ..., 132 of the n-th layer 110, ..., 113. The values ​​of the nodes 120, 121, 122 of the input layer 110 are equivalent to the input values ​​of the neural network 100, and the values ​​of the nodes 131 and 132 of the output layer 113 are equivalent to the output values ​​of the neural network 100. In addition, each edge 140, 141, 142 can have a real number, especially a real number in the interval [-1, 1] or the interval [0, 1] as a weight. Here, w (m,n) i,j represents the weight of the edge between the i-th node 120, ..., 132 of the m-th layer 110, ..., 113 and the j-th node 120, ..., 132 of the n-th layer 110, ..., 113. (n) i,j Defined as w (n,n+1) i,j .

[0130] In order to calculate the output value of the neural network, the input value is propagated through the neural network. The values ​​of the nodes 120, ..., 132 of the (n+1)th layer 110, ..., 113 can be calculated based on the values ​​of the nodes 120, ..., 132 of the nth layer 110, ..., 113 as follows:

[0131] x (n+1) j = f(Σ i x (n) i ∙ w (n) i,j)

[0132] Here, function f is a transfer function (also called an "activation function"). Known transfer functions are step functions, sigmoid functions (e.g., logistic function, generalized logistic function, hyperbolic tangent, inverse tangent function, error function, smoothed step function), or rectifier functions. Transfer functions are primarily used for normalization purposes.

[0133] The values ​​are also propagated layer by layer through the neural network, where the value of the input layer 110 is given by the input to the neural network 100, the value of the first hidden layer 111 can be calculated based on the value of the input layer 110 of the neural network 100, the value of the second hidden layer 112 can be calculated based on the value of the first hidden layer 111, and so on.

[0134] To determine the edge weight w (m、n) i,j , the neural network 100 must be trained by using training data. The training data particularly has training input data and training output data (in t i denoted by ). For a training step, the neural network 100 is applied to training input data to generate calculated output data. The training data and the calculated output data have, in particular, the same number of values ​​as the number of nodes of the output layer.

[0135] In particular, the comparison between the calculated output data and the training data is used to recursively adjust the weights in the neural network 100 (back propagation algorithm). In particular, the weights are modified according to the following formula:

[0136] w' (n) i,j = w (n) i,j – γ ∙ δ (n) j ∙ x (n) i

[0137] Where γ is the learning rate, and δ is the (n) j It can be calculated recursively as follows:

[0138] δ (n) j = (Σ k δ (n+1) k ∙ w (n+1) j,k ) ∙ f'(Σ i x (n) i ∙ w (n) i,j )

[0139] Based on δ (n+1) j , if the (n+1)th layer is not the output layer, and if the (n+1)th layer is the output layer 113, then

[0140] δ (n) j = (x (n+1) k -t (n+1) j ) ∙ f'(Σ i x (n) i ∙ w (n) i,j )

[0141] Where f' is the first-order derivative of the activation function, t (n+1) j is the comparative training value for the j-th node of the output layer 113.

[0142] Figure 3 A flow chart of a design solution of the method according to the present invention is shown. In a first step 31, two or more different operating ranges of the asynchronous motor are defined as a function of the slip rate and a function of the power supply frequency. In a second step 32 following the first step 31, at least one assigned anomaly detector is respectively assigned to each defined operating range, and the detector is suitable for detecting a fault state of the asynchronous motor within the corresponding operating range. In a third step 33 following the second step 32, the current operating range of the asynchronous motor is determined. In a fourth step 34 following the third step 33, at least one anomaly detector corresponding to the current operating range is activated, and the state of the asynchronous motor is monitored by the activated anomaly detector. In a fifth step 35 following the fourth step 34, a corresponding signal is generated when the activated anomaly detector detects an anomaly.

[0143] Figure 4 The definition of the different operating ranges of the asynchronous motor is shown. In step a), the value s of the slip s occurring during the operation of the asynchronous motor AM is first determined. i The slip value range S = [S min ; S max ]. Also determine the value f of the voltage supply frequency f occurring during operation of the asynchronous motor AM i The frequency range F = [f min ; f max In step b), the determined slip value range S is divided into Q slip sub-ranges S q The determined frequency value range is further divided into R frequency sub-ranges F rIn step c), Q×R operating ranges B(q, r) are defined, and the operating ranges are composed of the slip sub-ranges S q One of the frequency sub-ranges F r For this purpose, Q×R anomaly detectors A(q,r) are provided, wherein the anomaly detectors A(q,r) are assigned to the respective operating ranges B(q,r) for anomaly detection.

[0144] Figure 5 A design scheme in which multiple abnormality detectors are responsible for one operating range is shown. In the detection phase, after the current operating range B(q,r) of the asynchronous motor AM is determined, the abnormality detector A(q,r) corresponding to the current operating range B(q,r) is activated, and one or more abnormality detectors A(q-1,r-1), A(q-1,r), A(q-1,r+1), A(q,r-1), A(q,r+1), A(q+1,r-1), A(q+1,r), and A(q+1,r+1) corresponding to the operating ranges B(q-1,r-1), B(q-1,r), B(q-1,r+1), B(q,r-1), B(q,r+1), and B(q+1,r+1) adjacent to the current operating range B(q,r) of the asynchronous motor are activated. The state of the asynchronous motor AM is then monitored by the activated abnormality detectors A(q,r), A(q-1,r-1), A(q-1,r), A(q-1,r+1), A(q,r-1), A(q,r+1), A(q+1,r-1), A(q+1,r), A(q+1,r+1).

Claims

1. A method for condition monitoring of an asynchronous motor (AM), the asynchronous motor having a slip (s) and to which a supply voltage (U) having a supply frequency (f) is applied, the method comprising the following steps: - defining two or more different operating ranges (B) of an asynchronous motor (AM) as a function of the slip (s) and the supply frequency (f); - For each defined operating range (B), at least one associated anomaly detector (A i-1 、A i 、A i+1 ), the abnormality detector is adapted to detect a fault state (E) of the asynchronous motor (AM) in a corresponding operating range (B); - Confirm the current operating range (B) of the asynchronous motor (AM); - activating the at least one abnormality detector (A) correspondingly assigned to the current operating range (B), and monitoring the state of the asynchronous motor (AM) by means of the activated abnormality detector (A); as well as - In the event that an abnormality is detected by the activated abnormality detector (A), a corresponding signal is generated.

2. The method according to claim 1, comprising the steps of: - Determine the value of the slip (s) occurring during operation of an asynchronous motor (AM) i ) of the slip value range (S); - Determine the value of the mains frequency (f) occurring during operation of an asynchronous motor (AM) i )’s frequency range (F); - Divide the determined slip value range (S) into Q slip sub-ranges (S q ); - Divide the determined frequency value range into R frequency sub-ranges (F r ); - defining Q×R operating ranges (B(q,r)), wherein the operating range is defined by one of the slip subranges (Sq) and one of the frequency subranges (Fr); - Provide Q×R anomaly detectors (A(q,r)), where The anomaly detector (A(q, r)) is correspondingly assigned to the corresponding operating range (B(q, r)) to perform anomaly detection; - In the training phase, the anomaly detector (A(q,r)) is trained; - in a detection phase, determining a current operating range (B(q,r)) of the asynchronous motor (AM), activating an abnormality detector (A(q,r)) corresponding to the current operating range (B(q,r)), and monitoring the state of the asynchronous motor (AM) by means of the activated abnormality detector (A(q,r)); and - In case an anomaly is detected by the activated anomaly detector (A(q,r)), a signal indicative of said anomaly is generated by the activated anomaly detector (A(q,r)).

3. The method according to claim 2, comprising the steps of: - in a detection phase, after confirming a current operating range (B(q,r)) of the asynchronous motor (AM), activating an abnormality detector (A(q,r)) corresponding to the current operating range (B(q,r)), and activating one or more abnormality detectors (A(q-1,r-1), A(q-1,r), A(q-1,r+1), A(q,r-1), A(q,r+1), A(q+1,r-1), A(q+1,r), A(q+1,r+1)) corresponding to motor operating ranges adjacent to the current operating range (B(q,r)); and - The status of the asynchronous motor (AM) is monitored by the activated abnormality detectors (A(q,r), A(q-1,r-1), A(q-1,r), A(q-1,r+1), A(q,r-1), A(q,r+1), A(q+1,r-1), A(q+1,r), A(q+1,r+1)).

4. The method according to claim 1, comprising: - For each defined operating range (Bi), multiple anomaly detectors (A i-1 、A i 、A i+1 ),in, The corresponding anomaly detector (A i-1 、A i 、A i+1 )'s detection result exists as a result vector; - Apply majority decision on the result vector to determine whether an anomaly exists.

5. A method according to any one of the preceding claims, wherein The determination of the current operating range (B) of the asynchronous motor (AM) comprises the following steps: - Determine the value of slip (s) within the time series (T) i ) and the value of the power frequency (f) (f i ); - determining the current operating range (B) of the asynchronous motor (AM) based on the determined value of the slip (s) and the value of the power frequency (f).

6. The method according to claim 5, comprising the following steps: - Divide the time series (T) into ν subunits of shorter duration (T ν ),in, In each subunit, a measure for the slip (L s、ν ); - When the metric (L s、ν ) does not exceed the preset difference value, activate the anomaly detector (A), otherwise give up activating the anomaly detector (A).

Images