Multi-model fusion asynchronous motor fault diagnosis method

Abstract

The invention provides a multi-model fusion asynchronous motor fault diagnosis method. The method comprises the following steps: converting a collected asynchronous motor time domain signal into a frequency domain signal by adopting fast Fourier transform; extracting a plurality of characteristic values related to the motor fault from the frequency domain signal; the current working mode of the motor is determined by calculating the power frequency and the load rate of motor operation; using all normal samples to train a main model as a reference model; and for each working mode, training the sub-model by using the corresponding sample data. According to the working mode of the to-be-predicted sample, the corresponding sub-model is preferentially selected for calculation, and if the matched sub-model cannot be found, the main model is used for calculation. And calculating the fault probability of each component of the motor according to the abnormal probability of the characteristic value predicted and calculated by the model. The fault diagnosis method can effectively meet the fault diagnosis requirements in different working modes, is particularly suitable for fault diagnosis of the variable frequency motor under complex working conditions, and improves the accuracy and adaptability of diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of asynchronous motor fault diagnosis, and in particular to an asynchronous motor fault diagnosis method based on multi-model fusion. Background Art

[0002] With the rapid development of industrial automation, electric motors, as an important source of power, play a key role in various mechanical equipment. As an important component of modern industrial production, asynchronous motors are widely used in various mechanical transmission systems. Their normal operation is essential to ensure the safety and stability of the entire production line. In particular, asynchronous motors have been widely used in industrial production due to their simple structure, low cost and convenient maintenance. However, due to long-term operation in harsh working environments, asynchronous motors may have various faults, such as stator winding short circuit, rotor bar breakage, bearing wear, etc. These faults will not only lead to reduced motor efficiency and increased energy consumption, but may also cause safety accidents in severe cases, resulting in equipment damage or casualties. Therefore, effective fault diagnosis of asynchronous motors is of great significance. By timely discovering and handling potential faults, the reliability and safety of the motor system can be significantly improved, unplanned downtime can be reduced, and thus maintenance costs can be reduced and production efficiency can be improved.

[0003] At present, the asynchronous motor fault diagnosis methods mainly include methods based on signal analysis, methods based on physical models and methods based on artificial intelligence algorithms. These methods have solved the problem of motor fault diagnosis to a certain extent, but they also have the following limitations:

[0004] 1. Methods based on signal analysis: This type of method usually uses vibration signals, current signals, acoustic signals, etc. to identify the health status of the motor. For example, by analyzing the vibration spectrum characteristics of the motor when it is running, it can be determined whether there are problems such as imbalance, misalignment or bearing damage inside the motor; by analyzing the spectrum of the motor current signal, it can effectively detect faults such as stator winding short circuit and rotor bar breakage. This type of method is often limited by signal quality, making it difficult to detect early and weak signs of faults, and has weak recognition capabilities for complex composite faults, especially when multiple faults exist at the same time. It is difficult to accurately distinguish them. In addition, professional signal processing knowledge and technology are required, and the operation is relatively complicated for non-professionals.

[0005] 2. Methods based on physical models: This type of method is based on the physical model of the motor and diagnoses faults by comparing the difference between the actual measured data and the theoretical model prediction. For example, the ideal current waveform calculated using the motor equivalent circuit model is compared with the actual current waveform to detect abnormal conditions of the motor. The limitation of this method is that when faced with nonlinear systems or complex working conditions, the accuracy of the model will be affected, resulting in inaccurate diagnostic results. In addition, the establishment of a physical model requires precise motor parameters and good mathematical modeling capabilities, which are often difficult to achieve in practice.

[0006] 3. Methods based on artificial intelligence algorithms: In recent years, with the development of artificial intelligence technology, more and more researchers have begun to try to use machine learning algorithms such as neural networks, support vector machines, and decision trees to diagnose motor faults. These algorithms can automatically learn fault modes from large amounts of data to achieve more accurate fault identification. The limitation of this method is that it usually requires a large amount of high-quality labeled training data, which is a major challenge for practical applications. In addition, when faced with complex and changing working conditions, the generalization ability and adaptability of the model are still limited, especially under new working conditions, overfitting or underfitting may occur. In addition, it is difficult for non-professionals to understand and apply these algorithms.

[0007] 4. Diagnostic challenges under complex working conditions: Traditional fault diagnosis methods face greater challenges, especially under the complex working conditions of variable frequency motors. Variable frequency motors have diverse working modes, and the main frequency (power supply frequency) and load rate vary widely, which means that the fault characteristics will also change with the changes in these parameters.

[0008] In this case, it is difficult for a single fault diagnosis model to fully cover the changes in fault characteristics under all working conditions. Therefore, a diagnostic method that can adapt to multiple working conditions is needed to improve the accuracy and adaptability of diagnosis. Summary of the invention

[0009] In view of the above-mentioned problems, the purpose of the present invention is to provide an asynchronous motor fault diagnosis method based on multi-model fusion, so as to solve the limitations of traditional diagnosis methods under complex and changeable working conditions.

[0010] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0011] A multi-model fusion asynchronous motor fault diagnosis method comprises the following steps:

[0012] Fast Fourier transform is used to convert the collected asynchronous motor time domain signal into frequency domain signal;

[0013] Extract multiple characteristic values ​​related to motor faults from frequency domain signals to reflect the running characteristics of the motor under different fault conditions;

[0014] Determine the current working mode of the motor by calculating its operating frequency and load rate;

[0015] Use all normal samples to train a main model as a baseline model; for each working mode, use the corresponding sample data to train a sub-model;

[0016] According to the working mode of the sample to be predicted, the corresponding sub-model is selected for calculation. If no matching sub-model is found, the main model is used for calculation.

[0017] According to the abnormal probability of characteristic values ​​calculated by model prediction, the failure probability of each component of the motor is calculated.

[0018] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the extraction of multiple characteristic values ​​related to motor faults from frequency domain signals includes:

[0019] Extract eigenvalues ​​from the current spectrum, including: fuzzy interval eigenvalues, current harmonic eigenvalues, and Park eigenvalues;

[0020] Extract characteristic values ​​from the voltage spectrum, including voltage harmonic characteristic values ​​and unbalance.

[0021] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the calculation of the power frequency and load rate of the motor to determine its current working mode includes:

[0022] Determine the fundamental frequency: Use FFT to process the current signal to obtain frequency and amplitude data. According to the approximate range of the fundamental frequency, determine a search window. Based on the preset frequency range, find the frequency corresponding to the maximum amplitude in this search window, which is the fundamental frequency.

[0023] Calculate the load factor: For the sinusoidal current i(t) = Imsin(2πft), calculate the effective value of the current Irms as follows:

[0024]

[0025] Among them, Im is the maximum amplitude of the current, f is the frequency, and Im is the maximum amplitude of the sinusoidal current;

[0026] The load factor is calculated by dividing the effective current by the rated current.

[0027] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the LOF algorithm is used to train the main model, which specifically includes:

[0028] The local reachability density lrd(p) defines the density of a point p, which is calculated by considering the reachability distances of other points in the k-neighborhood of point p; for point p and any point o in its k-neighborhood, the reachability distance reach_dist(o, p) is defined as the following formula:

[0029]

[0030] Where k-dist(p) is the distance from point p to its kth nearest neighbor, d(o, p) is the Euclidean distance between point o and point p, and N k (p) is the set of points in the k-neighborhood of point p;

[0031] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the local reachable density lrd(p) of point p is defined as the following formula:

[0032]

[0033] Among them, |N k (p)| represents the number of points in the k-neighborhood of p.

[0034] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the local abnormality factor LOF(p) is used to describe the abnormality degree of point p relative to its neighborhood. For point p and any point o in its k-neighborhood, LOF(p) is defined as the following formula:

[0035]

[0036] Among them, the local outlier factor LOF(p) is the average value of the ratio of the average value of the local reachable density of all points in the k-neighborhood of point p to the local reachable density of point p itself;

[0037] The first H eigenvalues ​​are traversed separately, combined with the power frequency and load rate, and the LOF algorithm is used to train the model.

[0038] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the 3σ principle is used to train the main model, which specifically includes:

[0039] Assume that the random variable X follows a normal distribution N(μ, σ 2 ), where μ is the mean and σ is the standard deviation, then the mean (μ) is calculated as:

[0040]

[0041] Among them, xi is the eigenvalue of the i-th sample, and n is the total number of samples.

[0042] The formula for calculating the standard deviation (σ) is:

[0043]

[0044] For the last K eigenvalues, the mean and standard deviation of each eigenvalue sequence in the training sample are calculated, so the main model contains H LOF models and K statistical models.

[0045] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, the main model uses all training samples of normal motors, and the training samples of the sub-model are samples of a certain working mode whose number meets certain conditions;

[0046] When training the sub-model, first group all training samples according to the label of working mode and sum them. Assuming that there are N working modes whose occurrence times are greater than or equal to the total number of samples * 0.02, for these N working modes, use the corresponding samples to train the sub-model respectively. In the sub-model training stage, a total of N×H LOF models and N×K statistical models are saved;

[0047] The corresponding working mode is determined according to the load rate and fundamental frequency current of the sample to be predicted, and a check is made to see whether there is a sub-model matching the working mode; if so, the sub-model is used for prediction; if not, the main model is used for prediction instead.

[0048] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, for LOF model prediction:

[0049] For the first H eigenvalues, obtain the LOF model corresponding to its working mode according to whether there is a sub-model;

[0050] Use the LOF model to calculate the abnormal probability of the sample to be predicted;

[0051] For statistical model predictions:

[0052] For the last K eigenvalues, obtain the statistical model corresponding to its working mode according to whether there is a sub-model;

[0053] The standardized scores were calculated according to the 3σ rule, and the abnormal probability was calculated using a conversion function.

[0054] According to a multi-model fusion asynchronous motor fault diagnosis method provided by the present invention, when calculating the fault probability of each component of the motor, it specifically includes:

[0055] Feature weight assignment: For each type of fault, a weight coefficient is assigned to each relevant feature value, indicating the sensitivity of the feature value to the type of fault;

[0056] Comprehensive abnormality probability calculation: For each characteristic value, its abnormality probability Pi has been calculated. For each type of fault, its comprehensive abnormality probability Ptotal is calculated as follows:

[0057]

[0058] Where n is the number of eigenvalues ​​associated with this type of fault, wi is the weight coefficient of the i-th eigenvalue (the sum of all weight coefficients needs to be equal to 1), and Pi is the abnormal probability of the i-th eigenvalue;

[0059] Threshold setting: Set a threshold T. When the comprehensive abnormal probability of a certain type of fault exceeds the threshold, it is considered that the fault may exist.

[0060] It can be seen that compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The present invention can learn the changing trend of effective fault characteristics in motor signals in an unsupervised learning manner. By combining the prediction results of mathematical models and statistical models, the present invention improves the accuracy and reliability of prediction.

[0062] 2. By distinguishing between the main model and the sub-model, the present invention can flexibly select the most appropriate model for prediction according to different working modes. The use of the LOF model can effectively identify outliers, while the statistical model can capture the distribution characteristics of the data.

[0063] 3. The present invention completes motor fault diagnosis based on the idea of ​​abnormality detection, and identifies potential fault states by detecting abnormal features in the signal when the motor is running. This method based on abnormality detection can effectively identify potential faults of the motor under different working modes, thereby ensuring the safe operation of the motor system.

[0064] 4. The present invention determines the current working mode of the motor by calculating the operating frequency and load rate of the motor, and can classify the running state of the motor so as to select a more appropriate model for prediction later.

[0065] 5. The present invention uses all normal samples to train a main model as a benchmark model, and uses corresponding sample data to train sub-models for each working mode. These sub-models can more accurately reflect the operating status of the motor in different working modes.

[0066] 6. The present invention calculates the failure probability of each motor component based on the abnormal probability of the characteristic value calculated by the model prediction, and can convert the prediction result into a more intuitive failure probability to facilitate subsequent fault warning and diagnosis.

[0067] Therefore, the present invention can effectively cope with the fault diagnosis requirements under different working modes, and is particularly suitable for fault diagnosis of variable frequency motors under complex working conditions, thereby improving the accuracy and adaptability of diagnosis.

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of an embodiment of a multi-model fusion asynchronous motor fault diagnosis method of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0072] See also Figure 1 This embodiment provides a multi-model fusion asynchronous motor fault diagnosis method, which includes the following steps:

[0073] Step S1, using fast Fourier transform to convert the collected asynchronous motor time domain signal into a frequency domain signal.

[0074] Step S2, extracting multiple eigenvalues ​​related to motor faults from the frequency domain signal, a total of 55 eigenvalues ​​(41 current spectrum eigenvalues ​​and 14 voltage spectrum eigenvalues) to reflect the operating characteristics of the motor under different fault states.

[0075] Step S3, determining the current working mode of the motor by calculating the power frequency and load rate of the motor. Considering that the main frequency range is 10-80 Hz (the step size is 5, plus the left and right boundaries, a total of 16 values) and the load rate is 10%-120% (the step size is 10%, plus the left and right boundaries, a total of 14 values), this embodiment defines a total of 224 working modes.

[0076] Step S4, use all normal samples to train a main model as a benchmark model; for each working mode, use the corresponding sample data to train the sub-model. Among them, for each working mode, if the corresponding number of samples is greater than or equal to the total number of samples * 0.02, then train a sub-model of the corresponding working mode. During the training process, the local density anomaly factor (LOF) algorithm is used to train the model for the first 24 eigenvalues, and the 3σ rule is used to train the model for the last 31 eigenvalues.

[0077] Step S5, according to the working mode of the sample to be predicted, the corresponding sub-model is preferentially selected for calculation. If no matching sub-model is found, the main model is used for calculation.

[0078] Step S6, calculating the failure probability of each component of the motor in a weighted manner according to the abnormal probability of the characteristic value calculated by the model prediction.

[0079] In the above step S2, a variety of characteristic values ​​related to motor faults are extracted from the frequency domain signal, including:

[0080] 1. Extracting characteristic values ​​from current spectrum

[0081] (1) Fuzzy interval eigenvalue: between 0 and 600 Hz, with FL and RS as references, FL represents the power frequency, RS represents the rotation frequency, 12 intervals are divided from the current spectrum, and the maximum and mean values ​​of the 12 intervals are extracted respectively. Each interval corresponds to a eigenvalue that can reflect a specific fault type, which are counted as M1...M12 and A1...A12 respectively.

[0082] The calculation formulas for the maximum value Mi and the mean value Ai are as follows:

[0083] Mi=max(S(f)),fstart≤f≤fend

[0084]

[0085] Among them, f represents frequency, i represents the i-th fuzzy interval (the value range of i is 1...12), the frequency range of this interval is [fstart,fend], and the amplitude data corresponding to the frequency is S(f).

[0086] (2) Current harmonic characteristic value: Harmonic ratio: Calculate the ratio of the 2nd, 3rd, 5th, and 7th harmonics of the ABC three-phase current to the fundamental wave respectively.

[0087] Total Harmonic Distortion: Calculate the total harmonic distortion (THD) of the current, expressed as the following formula:

[0088]

[0089] Among them, I1 and In (the value range of n is 2...15) represent the effective values ​​of the fundamental current and each harmonic current respectively, and THD represents the total harmonic distortion rate calculated based on the effective value of the current.

[0090] Unbalance: Calculate the unbalance of the three-phase current, expressed as the following formula:

[0091]

[0092] (3) Park eigenvalue: By performing Park transform on the original ABC three-phase current signal, the D and Q two-phase currents are obtained, and other eigenvalues ​​related to motor faults are extracted through the FFT of the two-phase current vector modulus, such as stator turn-to-turn short circuit (P2), rotor fault (P1), and rotation frequency energy (P3). The Park transform of three-phase current (also called abc to dq transform) is a technique for converting a three-phase AC signal into a DC signal in a two-phase rotating coordinate system. This transform is very useful in the field of motor control and power electronics. The mathematical formula of the Park transform is as follows:

[0093]

[0094] Among them, ia, ib and ic represent the three-phase currents respectively, and α represents the angle of the rotating coordinate system relative to the stationary coordinate system.

[0095] The calculation formulas for P1, P2, and P3 are as follows:

[0096]

[0097] Among them, park_extend_amp represents the amplitude of Park vector modulus after FFT transformation, f represents frequency, FL represents power frequency, RS represents rotation frequency, park_dc represents the amplitude corresponding to the DC component, and PPF represents the pole passing frequency, which can be calculated by multiplying the slip by the number of poles. This characteristic frequency can be used to detect rotor broken bar faults.

[0098] 2. Extracting characteristic values ​​from voltage spectrum

[0099] (1) Voltage harmonic characteristic value:

[0100] Harmonic ratio: Calculate the ratio of the 2nd, 3rd, 5th and 7th harmonics of the ABC three-phase voltage to the fundamental wave.

[0101] Total harmonic distortion rate: Calculate the total harmonic distortion rate of voltage (the calculation method is the same as the total harmonic distortion rate of current).

[0102] (2) Unbalance: Calculate the unbalance of the three-phase voltage (the calculation method is the same as the current unbalance).

[0103] In the above step S3, the operating frequency and load rate of the motor are calculated to determine its current working mode, including:

[0104] (1) Determine the fundamental frequency:

[0105] Use FFT to process the current signal to obtain frequency and amplitude data. According to the approximate range of the fundamental frequency, determine a search window, and search for the frequency corresponding to the maximum amplitude in this search window based on the preset frequency range, which is the fundamental frequency. The preset frequency range of the fundamental frequency search window is 10Hz to 80Hz.

[0106] (2) Calculation of load factor: The calculation of load factor depends on the effective value of current Irms and the rated current. The load factor can be calculated by dividing the effective value of current by the rated current. For a sinusoidal current i(t) = Imsin(2πft), the effective value of current Irms can be expressed as follows:

[0107]

[0108] Among them, Im is the maximum amplitude of the current, f is the frequency, and Im is the maximum amplitude of the sinusoidal current;

[0109] In the above step S4, the LOF algorithm is used to train the main model, which specifically includes:

[0110] The Local Outlier Factor (LOF) algorithm is an unsupervised learning method for detecting outliers in a dataset. The core concept of the LOF algorithm is to determine whether a point is an outlier based on the concept of density. In this algorithm, the local reachable density (LRD) and the local outlier factor (LOF) are two key components. In motor fault diagnosis, the eigenvalue of each sample (the eigenvector of the motor when it is running) is assigned a LOF score. The higher the score, the more likely the sample is to represent a potential fault state. LOF does not require pre-labeled anomaly data, which is very important in practical applications because labeling anomaly data is usually difficult and expensive.

[0111] The local reachability density lrd(p) defines the density of a point p, which is calculated by considering the reachability distances of other points in the k-neighborhood of point p; for point p and any point o in its k-neighborhood, the reachability distance reach_dist(o, p) is defined as the following formula:

[0112]

[0113] Where k-dist(p) is the distance from point p to its kth nearest neighbor, d(o, p) is the Euclidean distance between point o and point p, and N k (p) is the set of points in the k-neighborhood of point p;

[0114] The local reachability density lrd(p) of a point p is defined as follows:

[0115]

[0116] Among them, |N k (p)| represents the number of points in the k-neighborhood of p.

[0117] The local outlier factor LOF(p) is used to describe the degree of outlier of point p relative to its neighborhood. For point p and any point 0 in its k-neighborhood, LOF(p) is defined as the following formula:

[0118]

[0119] The local outlier factor LOF(p) is the average of the ratios of the average of the local reachable densities of all points in the k-neighborhood of point p to the local reachable density of point p itself.

[0120] Based on the above-mentioned calculation formula, the first H eigenvalues ​​are traversed respectively, combined with the power frequency and load rate, and the LOF algorithm is used to train the model. Among them, the H value of this embodiment is 24.

[0121] In statistics, the 3σ principle is a method used to evaluate data distribution, especially in the case of normal distribution. The 3σ principle is based on the fact that for a normally distributed data set, approximately 99.73% of the data values ​​fall within 3 standard deviations (σ) of the mean (μ). This means that the vast majority of observations will fall between μ-3σ and μ+3σ. In motor fault diagnosis, since the extracted motor eigenvalue data usually presents an approximate normal distribution characteristic, the use of the 3σ principle can effectively identify abnormal values ​​outside the normal range. For the data distribution of each eigenvalue, if the eigenvalue in the monitoring phase is outside the 3σ range of the learning phase, it can be considered that the abnormal probability of this eigenvalue is high, which helps to accurately identify potential fault conditions.

[0122] Assume that the random variable X follows a normal distribution N(μ, σ 2 ), where μ is the mean and σ is the standard deviation, then the mean (μ) is calculated as:

[0123]

[0124] Among them, xi is the eigenvalue of the i-th sample, and n is the total number of samples.

[0125] The formula for calculating the standard deviation (σ) is:

[0126]

[0127] Based on the above-mentioned related calculation formulas, for the last K eigenvalues, the mean and standard deviation of each eigenvalue sequence in the training sample are calculated, so the main model includes H LθF models and K statistical models. Among them, the K value in this embodiment is 31.

[0128] In this embodiment, the training process of the sub-model is similar to that of the main model, but the training samples used are different. The main model uses all the training samples of the normal motor, while the training samples of the sub-model are samples of a certain working mode whose number meets certain conditions.

[0129] When training the sub-model, first group all training samples according to the working mode label and sum them up. Assuming that there are N working modes whose occurrence times are greater than or equal to the total number of samples * 0.02, for these N working modes, use the corresponding samples to train the sub-model respectively. In the sub-model training stage, a total of N×24 LOF models and N×31 statistical models are saved.

[0130] Therefore, both the main model and the sub-model in this embodiment use the LOF algorithm to detect abnormal changes in the first 24 eigenvalues, while the statistical model calculates the mean and standard deviation of the last 31 eigenvalues, and then uses the 3σ rule to calculate the abnormal changes in the last 31 eigenvalues ​​during the monitoring phase. The main model is trained using all sample data, while the sub-model is trained on sample data for a specific operating mode. These models can be used to calculate the state of the motor and potential abnormalities during the monitoring phase. By combining the LOF algorithm and the statistical model, the optimal model is selected under different operating modes, so that the system can not only adapt to different operating modes, but also provide more accurate fault diagnosis.

[0131] In the above step S5, the corresponding working mode is determined according to the load rate and fundamental frequency current of the sample to be predicted, and it is checked whether there is a sub-model matching the working mode; if so, the sub-model is used for prediction; if not, the main model is used for prediction instead.

[0132] In this embodiment, the LOF model predicts:

[0133] For the first 24 eigenvalues, the LOF model corresponding to its working mode is obtained according to whether there is a sub-model; the LOF model is used to calculate the abnormal probability of the sample to be predicted. In the LOF algorithm, if a certain eigenvalue is relatively small, then in theory its abnormal probability is relatively high. However, in actual spectrum analysis, usually only the eigenvalues ​​with higher spectral components are concerned, and those smaller eigenvalues ​​are often ignored. Therefore, if the current eigenvalue is less than the mean plus 1 standard deviation in the learning phase, the abnormal probability is set to 0.01. This can avoid the high abnormal probability caused by too small eigenvalues, thereby ensuring that the calculation of abnormal probability is more reasonable and accurate. In this way, the potential faults of the motor in different working modes can be effectively identified, while avoiding misjudgments caused by too small eigenvalues, ensuring the safe operation of the motor system.

[0134] In this embodiment, for the statistical model prediction:

[0135] For the last 31 eigenvalues, the statistical model corresponding to its working mode is obtained according to whether a sub-model exists; the standardized score is calculated according to the 3σ rule, and the abnormal probability is calculated using the conversion function.

[0136] Therefore, this embodiment can flexibly select the most appropriate model for prediction according to different working modes by distinguishing between the main model and the sub-model. The LOF model can effectively identify outliers, while the statistical model can capture the distribution characteristics of the data. The prediction results of the mathematical model and the statistical model are combined to improve the accuracy and reliability of the prediction.

[0137] In the above step S5, the abnormal probability of each of the 55 eigenvalues ​​has been calculated. In actual operation, the motor fault types are mainly divided into the following six categories: stator fault, rotor fault, air gap fault, bearing fault, transmission system fault, and loose foundation. Since different fault types have different sensitivities to eigenvalues, each type of fault will have an eigenvalue that has a greater impact on it. For example, the main eigenvalues ​​that affect stator faults are IT2, P2, and M7; the main eigenvalues ​​that affect bearing faults are A8, A10, M8, and M10. In order to more accurately evaluate the abnormal probability of each component of the motor, the following steps are taken:

[0138] Feature weight assignment: For each type of fault, a weight coefficient is assigned to each relevant feature value based on historical data and expert experience, indicating the sensitivity of the feature value to the type of fault. For example, the weights of stator faults IT2, P1, and M7 are wIT2, wP1, and wM7, respectively.

[0139] Comprehensive abnormality probability calculation: For each characteristic value, its abnormality probability Pi has been calculated. For each type of fault, its comprehensive abnormality probability Ptotal is calculated as follows:

[0140]

[0141] Among them, n is the number of eigenvalues ​​related to this type of fault, wi is the weight coefficient of the i-th eigenvalue (the sum of all weight coefficients needs to be equal to 1), and Pi is the abnormal probability of the i-th eigenvalue.

[0142] Threshold setting: A threshold T is set based on actual experience and historical data. When the comprehensive abnormal probability of a certain type of fault exceeds the threshold, it is considered that this type of fault may exist.

[0143] Through the above steps, this embodiment can effectively evaluate the abnormal probability of each component of the motor, thereby achieving accurate fault diagnosis. This method can not only improve the accuracy and reliability of diagnosis, but also better meet the fault diagnosis needs of the motor in different working modes.

[0144] In summary, the present invention can learn the changing trend of effective fault features in motor signals in an unsupervised learning manner. By combining the prediction results of mathematical models and statistical models, the present invention improves the accuracy and reliability of prediction. By distinguishing between the main model and the sub-model, the present invention can flexibly select the most appropriate model for prediction according to different working modes. The use of the LOF model can effectively identify abnormal points, while the statistical model can capture the distribution characteristics of the data. Therefore, the present invention can effectively respond to the fault diagnosis needs under different working modes, and is particularly suitable for fault diagnosis of variable frequency motors under complex working conditions, improving the accuracy and adaptability of diagnosis.

[0145] Furthermore, the present invention completes motor fault diagnosis based on the idea of ​​abnormality detection, and identifies potential fault conditions by detecting abnormal features in the signal when the motor is running. This method based on abnormality detection can effectively identify potential faults of the motor under different operating modes, thereby ensuring the safe operation of the motor system.

[0146] Furthermore, the present invention determines the current working mode of the motor by calculating the operating frequency and load rate of the motor, and can classify the running state of the motor so as to select a more appropriate model for prediction later.

[0147] Furthermore, the present invention uses all normal samples to train a main model as a benchmark model, and uses corresponding sample data to train sub-models for each operating mode. These sub-models can more accurately reflect the operating status of the motor in different operating modes.

[0148] Furthermore, the present invention calculates the failure probability of each motor component based on the abnormal probability of the characteristic value calculated by the model prediction, and can convert the prediction result into a more intuitive failure probability to facilitate subsequent fault warning and diagnosis.

[0149] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0150] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A multi-model fusion asynchronous motor fault diagnosis method, characterized in that: The following steps are involved: Fast Fourier transform is used to convert the collected asynchronous motor time domain signal into frequency domain signal; Extract multiple characteristic values ​​related to motor faults from frequency domain signals to reflect the running characteristics of the motor under different fault conditions; Determine the current working mode of the motor by calculating its operating frequency and load rate; Use all normal samples to train a main model as a baseline model; for each working mode, use the corresponding sample data to train a sub-model; According to the working mode of the sample to be predicted, the corresponding sub-model is selected for calculation. If no matching sub-model is found, the main model is used for calculation; According to the abnormal probability of characteristic values ​​calculated by model prediction, the failure probability of each component of the motor is calculated.

2. The method according to claim 1, characterized in that: The method extracts multiple characteristic values ​​related to motor faults from the frequency domain signal, including: Extract eigenvalues ​​from the current spectrum, including: fuzzy interval eigenvalues, current harmonic eigenvalues, and Park eigenvalues; Extract characteristic values ​​from the voltage spectrum, including voltage harmonic characteristic values ​​and unbalance.

3. The method according to claim 1, characterized in that The calculating of the power frequency and load rate of the motor to determine its current working mode includes: Determine the fundamental frequency: Use FFT to process the current signal to obtain frequency and amplitude data. According to the approximate range of the fundamental frequency, determine a search window. Based on the preset frequency range, find the frequency corresponding to the maximum amplitude in this search window, which is the fundamental frequency. Calculate the load factor: For the sinusoidal current i(t) = Imsin(2πft), calculate the effective value of the current Irms as follows: Among them, Im is the maximum amplitude of the current, f is the frequency, and Im is the maximum amplitude of the sinusoidal current; The load factor is calculated by dividing the effective current by the rated current.

4. The method according to claim 1, characterized in that: Use the LOF algorithm to train the main model, including: The local reachability density lrd(p) defines the density of a point p, which is calculated by considering the reachability distances of other points in the k-neighborhood of point p; for point p and any point o in its k-neighborhood, the reachability distance reach_dist(o, p) is defined as the following formula: Where k-dist(p) is the distance from point p to its kth nearest neighbor, d(o, p) is the Euclidean distance between point o and point p, and N k (p) is the set of points in the k-neighborhood of point p.

5. The method according to claim 4, characterized in that: The local reachability density lrd(p) of a point p is defined as follows: Among them, |N k (p)| represents the number of points in the k-neighborhood of p.

6. The method according to claim 5, characterized in that: The local outlier factor LOF(p) is used to describe the degree of outlier of point p relative to its neighborhood. For point p and any point o in its k-neighborhood, LOF(p) is defined as the following formula: Among them, the local outlier factor LPF(p) is the average value of the ratio of the average value of the local reachable density of all points in the k-neighborhood of point p to the local reachable density of point p itself; The first H eigenvalues ​​are traversed separately, combined with the power frequency and load rate, and the LOF algorithm is used to train the model.

7. The method according to claim 6, characterized in that: Use the 3σ principle to train the main model, including: Assume that the random variable X follows a normal distribution N(μ, σ 2 ), where μ is the mean and σ is the standard deviation, then the mean (μ) is calculated as: Among them, xi is the eigenvalue of the i-th sample, and n is the total number of samples. The formula for calculating the standard deviation (σ) is: For the last K eigenvalues, the mean and standard deviation of each eigenvalue sequence in the training sample are calculated, so the main model contains H LOF models and K statistical models.

8. The method according to claim 7, characterized in that: The main model uses all training samples of normal motors, while the training samples of the sub-model are samples of a certain working mode whose quantity meets certain conditions; When training the sub-model, first group all training samples according to the label of working mode and sum them. Assuming that there are N working modes whose occurrence times are greater than or equal to the total number of samples * 0.02, for these N working modes, use the corresponding samples to train the sub-model respectively. In the sub-model training stage, a total of N×H LOF models and N×K statistical models are saved; Determine the corresponding working mode according to the load rate and fundamental frequency current of the sample to be predicted, and check whether there is a sub-model matching the working mode; if so, use the sub-model for prediction; If it does not exist, the main model is used for prediction.

9. The method according to claim 7, characterized in that: For LOF model prediction: For the first H eigenvalues, obtain the LOF model corresponding to its working mode according to whether there is a sub-model; Use the LOF model to calculate the abnormal probability of the sample to be predicted; For statistical model predictions: For the last K eigenvalues, obtain the statistical model corresponding to its working mode according to whether there is a sub-model; The standardized scores were calculated according to the 3σ rule, and the abnormal probability was calculated using a conversion function.

10. The method according to any one of claims 1 to 9, characterized in that: When calculating the failure probability of each motor component, specifically include: Feature weight assignment: For each type of fault, a weight coefficient is assigned to each relevant feature value, indicating the sensitivity of the feature value to the type of fault; Comprehensive abnormality probability calculation: For each characteristic value, its abnormality probability Pi has been calculated. For each type of fault, its comprehensive abnormality probability Ptotal is calculated as follows: Where n is the number of eigenvalues ​​associated with this type of fault, wi is the weight coefficient of the i-th eigenvalue, and Pi is the abnormal probability of the i-th eigenvalue; Threshold setting: Set a threshold T. When the comprehensive abnormal probability of a certain type of fault exceeds the threshold, it is considered that the fault may exist.

Images