A motor compound fault identification method based on adaptive anomaly detection

Abstract

The present invention discloses a motor composite fault identification method based on adaptive anomaly detection, the method comprising collecting motor high-frequency current data for preprocessing to form a sample data set; for small sample data, learning Gaussian distribution key parameters based on the starfinch optimization algorithm, establishing a Gaussian distribution anomaly identification model, using the k-sigma method to learn the threshold, and screening abnormal data; for large sample data, training the crow search optimization stacked autoencoder with normal data, and screening abnormal data by whether the original data can be reconstructed; for abnormal data, based on source domain and target domain samples, a coordinated multi-source domain adaptive method is used to extract the invariant features of the fault sample domain to achieve fault classification; for normal data, piecewise linear fitting is performed, and the time when the sub-healthy equipment fails is predicted according to the spliced ​​trend fitting line parameters. The present invention can identify complex situations with multiple concurrent faults and achieve accurate fault prediction.

Description

Technical Field

[0001] The invention belongs to the field of motor compound fault diagnosis, and in particular relates to a motor compound fault identification method based on adaptive anomaly detection. Background Art

[0002] With the rapid development of industrial technology, motors, as core power equipment, are increasingly used in various fields. However, motors are susceptible to various faults during long-term operation. Some of these faults are single, and some are multiple faults coexisting, which brings great challenges to motor fault diagnosis.

[0003] Faults come from different components inside the motor, such as windings, bearings, and insulation layers, and from the external environment, such as temperature, humidity, and load changes. During their development, these faults couple and stimulate each other, causing the fault characteristics to become more complicated and diversified, increasing the difficulty of diagnosing complex motor faults. Summary of the invention

[0004] The purpose of the present invention is to cope with the complex situation of multiple concurrent faults and provide a motor composite fault identification method based on adaptive abnormality detection, which can effectively improve the reliability of the motor, timely and accurately diagnose the composite faults in the motor, and increase the service life of the motor.

[0005] In order to solve the above technical problems, the present invention provides a motor compound fault identification method based on adaptive abnormality detection, comprising:

[0006] The zero-crossing smoothing outlier method is used to preprocess the motor high-frequency current data to obtain a sample data set for motor composite fault diagnosis.

[0007] For small sample data, the key parameters of Gaussian distribution are learned based on the Starfinch optimization algorithm, a Gaussian distribution anomaly recognition model is established, and the k-sigma method is used to learn the threshold and filter abnormal data;

[0008] For large sample data, we use normal data to train the crow search optimization stacked autoencoder, and filter out abnormal data by whether the original data can be reconstructed;

[0009] For abnormal data, a collaborative multi-source domain adaptive method is used to extract the invariant features of the fault sample domain to achieve fault classification;

[0010] For normal data, piecewise linear fitting is performed, and the time when sub-healthy equipment will fail is predicted based on the spliced ​​trend fitting line parameters.

[0011] Optionally, the preprocessing of the motor high-frequency current data by using the zero-crossing smoothing outlier method includes:

[0012] During the operation of the motor, when rotor imbalance and three-phase imbalance, stator-rotor misalignment and rotor imbalance, stator-rotor misalignment and overload compound faults occur, high-frequency current data is collected using a high-voltage resistant high-precision current transformer;

[0013] The collected high-frequency current data is subjected to zero-crossing detection, outlier detection and data smoothing to remove abnormal and redundant data, reduce data noise, make the data smoother and easier to analyze, and form a sample data set for motor composite fault diagnosis after preprocessing.

[0014] Optionally, the adopting a Gaussian distribution anomaly recognition model to perform abnormal data recognition includes:

[0015] For small sample data, the starfinch optimization algorithm is used to learn the key parameters of the Gaussian distribution from the data, and a Gaussian distribution anomaly recognition model is established. The k-sigma method is used to learn the normal probability threshold, and the probability value of each data point under the Gaussian distribution is calculated. The proportion of the probability value of each data point in the statistical segment that is less than the normal probability threshold is counted, and the IQR method is used to determine the proportion threshold of the normal data segment. When the proportion exceeds the proportion threshold, it is identified as abnormal data.

[0016] Optionally, the method of using a crow search optimized stacked autoencoder reconstruction method to realize abnormal data recognition includes:

[0017] For large sample data, normal data is collected as training data to ensure that the data set contains sufficient normal samples to capture its characteristic distribution. The normal data is used to train the crow search optimized stacked autoencoder. The crow search optimized stacked autoencoder learns the reconstruction error distribution of normal data and obtains the reconstruction error threshold of normal data. The data to be detected is input into the trained autoencoder to calculate the reconstruction error of the data to be detected. When the reconstruction error exceeds the normal threshold, it is identified as abnormal data.

[0018] Optionally, the method of extracting invariant features of fault sample domains by using a coordinated multi-source domain adaptive method to implement fault classification includes:

[0019] Collect samples from abnormal data source domain and target domain, adopt a collaborative multi-source domain adaptive method to extract invariant features of fault sample domain, use the extracted invariant features of fault sample domain to train the classifier, and apply the trained classifier to the target domain data to realize fault classification of samples.

[0020] Optionally, predicting the time of failure of the sub-health equipment according to the spliced ​​trend fitting line parameters includes:

[0021] Current characteristics are extracted from high-frequency current data, and the characteristic data are divided into multiple segments. A linear function is used to fit the current characteristic values ​​of each segment to obtain the changing trend line of the current characteristics. The sparrow search algorithm is used to optimize the slope and intercept of the trend line. The trend lines of each segment are spliced ​​in sequence from beginning to end to form a trend fitting line of data splicing. The quantum genetic algorithm is used to obtain the warning threshold of the current characteristic amplitude. The time when sub-healthy equipment will fail is predicted based on the three parameters of the slope, intercept and warning threshold of the current characteristic amplitude of the trend fitting line.

[0022] The motor composite fault identification method provided by the present invention comprises collecting high-frequency current data of the motor for preprocessing to form a sample data set; for small sample data, learning the key parameters of Gaussian distribution based on the starfinch optimization algorithm, establishing a Gaussian distribution abnormal recognition model, using the k-sigma method to learn the threshold, and filtering abnormal data; for large sample data, training the crow search optimization stacked autoencoder with normal data, and filtering abnormal data by whether the original data can be reconstructed; for abnormal data, extracting the invariant features of the fault sample domain based on the source domain and target domain samples using a coordinated multi-source domain adaptive method to achieve fault classification; for normal data, performing piecewise linear fitting, and predicting the time when the sub-health equipment fails according to the spliced ​​trend fitting line parameters. The present invention can identify complex situations with multiple concurrent faults and achieve accurate fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 A schematic diagram of a flow chart of a motor composite fault identification method provided by an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of a Starfinch optimization algorithm flow chart of a motor composite fault identification method provided by an embodiment of the present invention;

[0026] Figure 3 A schematic diagram of a stacked autoencoder framework for a motor composite fault identification method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] like Figure 1 As shown, Figure 1 The present invention provides a flow chart of a motor composite fault identification method according to an embodiment of the present invention. The motor composite fault identification method specifically includes five contents.

[0029] S11: Preprocess the high-frequency current data of the motor using the zero-crossing smoothing outlier method to obtain a sample data set for motor composite fault diagnosis.

[0030] It should be noted that in order to collect high-frequency current data, the open-and-close Hall current transformer is connected to the power cord of the distribution cabinet, and the original stator current data of the motor working for a long time is collected. The original data is segmented and sliced ​​according to the time length of 10 seconds to obtain sliced ​​data. The zero-crossing smoothing outlier method is used to remove abnormal and redundant data from the sliced ​​data to reduce data noise. It specifically includes three steps.

[0031] S111: Perform zero-crossing judgment on the sinusoidal current, set a zero-crossing number threshold, judge the number of times the current signal in a slice crosses the zero point, and mark the current data segment with a number less than the threshold as abnormal.

[0032] S112: Use the heterogeneity ratio algorithm to screen outliers in the current data, calculate the heterogeneity ratio between data points based on the average distance of the k-neighborhood of the sliced ​​data points, and obtain a heterogeneity ratio matrix. Construct a heterogeneity ratio factor by calculating the inner product of each row element in the heterogeneity ratio matrix, sort the values ​​of the heterogeneity ratio factors of all data points, set the factor threshold, and delete the current data points exceeding the threshold as outliers.

[0033] S113: Use moving average smoothing method to reduce data noise. According to the characteristics of the data and the smoothing requirements, select an appropriate smoothing window size, calculate the average value of the data points in the window, move the smoothing window one position to the right, repeatedly calculate the moving average value, and replace the corresponding data points of the original data with the moving average value.

[0034] Through the above method, zero-crossing judgment is realized, outlier data is eliminated, and data noise is reduced. After preprocessing, a sample data set for motor composite fault diagnosis is formed.

[0035] Based on the above discussion, in an optional embodiment of the present application, the data preprocessing setting process using the zero-crossing smoothing outlier method specifically includes:

[0036] In an optional embodiment of S11, in the preprocessing stage, the threshold is set to 90 percent of the number of zero crossings within 10 seconds that is less than the frequency of the current signal; the neighbor parameter is set to 28, the factor threshold is set to 7950; and the smoothing window size is set to 5 data points.

[0037] S12: For small sample data, the key parameters of Gaussian distribution are learned based on the starfinch optimization algorithm, a Gaussian distribution anomaly recognition model is established, and the k-sigma method is used to learn the threshold and filter abnormal data.

[0038] It should be noted that the Gaussian distribution anomaly recognition model is introduced in this application in order to screen abnormal data and discover potential fault problems early, so as to take corresponding measures to deal with them and improve the stability and reliability of the motor.

[0039] The starfinch optimization algorithm (NOA) is a meta-heuristic algorithm based on nature inspiration. It combines the characteristics of the particle swarm optimization algorithm and the artificial bee colony algorithm. It is inspired by the foraging, storage, cache search and recovery behaviors of starfinch. The NOA algorithm is applied to the optimization problem to learn the key parameters of the Gaussian distribution, such as Figure 2 As shown, it specifically includes four steps.

[0040] S121: Initialize the population and randomly generate a group of star finch individuals, each of which represents the mean and variance of each Gaussian distribution.

[0041] S122: Calculate the fitness value of each individual, and select excellent individuals as storage points based on the fitness. These individuals are the positions of the current population, that is, the mean and variance.

[0042] S123: Simulate the behavior of the starfinch using spatial memory to retrieve stored food, and update the position of the individual by calculating the value of the objective function, that is, update the mean and variance. The objective function is a negative log-likelihood function based on the sample data, and its formula is:

[0043]

[0044] In the formula, c is the number of samples, X i is the i-th current value of the current data, μ is the mean, σ 2 is the variance.

[0045] S124: When the maximum number of iterations is reached, the optimal solution obtained is the key parameter of the Gaussian distribution.

[0046] The k-sigma method is used to learn the normal probability threshold. The specific implementation process of the k-sigma method is as follows:

[0047] 1. According to the Gaussian distribution anomaly recognition model, the mean μ and standard deviation σ of the current data are obtained;

[0048] 2. Choose a suitable p value based on actual needs and data characteristics;

[0049] 3. Through the p value, the upper control limit of the normal probability threshold is μ+pσ, and the lower control limit is μ-pσ.

[0050] The IQR method is used to determine the ratio threshold of the normal current data segment and filter out abnormal data. The specific implementation process of the IQR method is as follows:

[0051] 1. Sort all data points in the normal current data segment from small to large;

[0052] 2. Find the 25% position in the current data as E1, and the 75% position in the current data as E3;

[0053] 3. IQR = E3-E1;

[0054] 4. Determine the proportional threshold range of the normal high-frequency current data segment, with the upper limit being E3+3IQR and the lower limit being E1-3IQR;

[0055] 5. Finally, filter out abnormal data.

[0056] Through the above method, a Gaussian distribution anomaly recognition model is established, the probability value of each data point under the Gaussian distribution is calculated, the proportion of the probability value of each data point in the current data segment that is less than the normal probability threshold is counted, the proportion threshold of the normal current data segment is determined, and abnormal data is screened.

[0057] Based on the above discussion, in an optional embodiment of the present application, the process of setting abnormal data identification using the Gaussian distribution abnormality identification model specifically includes:

[0058] In an optional embodiment of S12, the k-sigma method is used to learn the normal probability threshold, and the p value in the k-sigma method is set to 3, so that the upper control limit of the normal probability threshold is μ+3σ, and the lower control limit is μ-3σ.

[0059] S13: For large sample data, use normal data to train crow search to optimize stacked autoencoders, and filter out abnormal data by whether the original data can be reconstructed.

[0060] It should be noted that this application introduces the crow search optimization stacked autoencoder reconstruction method, uses normal current data to train each layer of autoencoder, and combines the crow search algorithm to optimize the mapping direction of fault classification, extract effective fault classification feature information for reconstruction, and screen abnormal data, such as Figure 3 As shown, it specifically includes three steps.

[0061] S131: Use normal current data to train the first autoencoder to obtain the hidden layer output h 1 ,h 1 The second autoencoder is trained as the input of the next layer, and the above process is repeated until the stacked autoencoder is built. The functional relationship of the stacked autoencoder is:

[0062]

[0063]

[0064] Where W e is the weight matrix of the encoder, b e is the bias vector of the encoder, x is the normal current data, h is the hidden layer output of the stacked autoencoder, f is the hidden layer of the stacked autoencoder, n is the total number of hidden layers of the stacked autoencoder, W d is the decoder weight matrix, b d is the decoder bias vector, and y is the reconstructed data of the decoder.

[0065] S132: The stacked autoencoder is optimized and trained using the crow search algorithm. First, the crow population is initialized and the fitness is calculated. Then, the tracking object, the position and the historical optimal solution are iteratively selected. After the termination condition is met, the parameters corresponding to the optimal crow individual are selected as the optimized stacked autoencoder parameters.

[0066] S133: The stacked autoencoder obtains a reconstruction error threshold of normal current data by learning the reconstruction error distribution of normal current data, and inputs the current data to be detected into the trained stacked autoencoder for reconstruction. When the reconstruction error of the current data to be detected exceeds the normal threshold, it is identified as abnormal data.

[0067] Based on the above discussion, in an optional embodiment of the present application, the process of setting up the above-mentioned abnormal data screening using the crow search optimized stacked autoencoder reconstruction method specifically includes:

[0068] In an optional embodiment of S13, the encoder structure is 52-33-21-17-11, the decoder structure is 11-17-21-33-52, the stacked autoencoder training batch size is 32, and iterative fine-tuning is performed 20 times.

[0069] S14: For abnormal data, a collaborative multi-source domain adaptive method is used based on source domain and target domain samples to extract fault sample domain invariant features to achieve fault classification.

[0070] It should be noted that abnormal data includes abnormal operating conditions and classified faults. Taking abnormal operating conditions as a type of classified faults, this application introduces a collaborative multi-source domain adaptive method. Through the cooperation of independent source domain models and independent target domain models, the invariant features of the fault domain are extracted to reduce the impact of domain differences and ultimately improve the accuracy of fault classification diagnosis. It specifically includes three steps.

[0071] S141: Use abnormal samples from the source domain to train an independent source domain model M S , independent source domain model M S It consists of a feature extractor and a classifier, using the cross entropy loss function And the regularized loss function Train an independent source domain model M S , upload to the target domain.

[0072] S142: Regularized loss function using high confidence data in abnormal samples of the target domain And the cross entropy loss function for low confidence data Add support for independent target domain model M T The recognition rate of the aggregated independent source domain model M S and independent target domain model M T Get the overall model M G。

[0073] S143: The trained overall model M G Applied to abnormal samples in the target domain, the invariant features of the fault sample domain are extracted to train the classifier and realize the fault classification of the samples.

[0074] Based on the above discussion, in an optional embodiment of the present application, the setting process of the above-mentioned multi-source domain adaptation method using collaborative cooperation specifically includes:

[0075] In an optional embodiment of S14, when setting the collaborative multi-source domain adaptation method, the confidence threshold is 0.9, SGD is selected as the optimization algorithm, and the batch size is 40.

[0076] S15: Perform piecewise linear fitting on normal data and predict the time when sub-health equipment will fail based on the spliced ​​trend fitting line parameters.

[0077] It should be noted that time domain and frequency domain features are extracted from high-frequency current data, and the values ​​of some of these features continue to rise as the health status of the equipment deteriorates. The rising trend fitting line of these features represents the degree of degradation of the health status of the equipment. The sparrow search algorithm is used to optimize the slope and intercept of the trend line. The sparrow search algorithm has the advantages of strong optimization ability and fast convergence speed, which improves the prediction efficiency of fault diagnosis. The quantum genetic algorithm is used to obtain the warning threshold of the current characteristic amplitude. The time when the sub-healthy equipment will fail is predicted according to the parameters of the trend fitting line and the warning threshold of the current characteristic amplitude. It specifically includes three steps.

[0078] S151: Collect current characteristic data, divide the characteristic data into multiple segments, fit each segment of current characteristic value with a linear function, and obtain a change trend line of the current characteristic.

[0079] S152: Using the sparrow search algorithm to find the slope and intercept of the trend line, the trend line of each section is spliced ​​in sequence from beginning to end to form a trend fitting line of data splicing.

[0080] S153: A quantum genetic algorithm is used to obtain a warning threshold of the current characteristic amplitude. Combined with the slope and intercept of the trend line, when the current characteristic trend line of the device to be detected reaches the warning threshold from the amplitude prediction point, the device will fail, and the time to failure is calculated. The functional relationship is:

[0081]

[0082] Where v is the slope, which indicates the rate at which the current characteristic data deteriorates over time; r is the intercept, which indicates the starting point of the trend fitting line; and q is the warning threshold of the current characteristic amplitude.

[0083] Based on the above discussion, in an optional embodiment of the present invention, the time setting process for predicting the occurrence of a failure of the above-mentioned sub-health device specifically includes:

[0084] In an optional embodiment of S15, the maximum number of iterations of the sparrow search algorithm is set to 100, the proportion of discoverers is 20%, the proportion of investigators is 10%, and the safety threshold is 0.7.

[0085] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Specific examples are used herein to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A motor compound fault identification method based on adaptive anomaly detection, characterized in that: include: The zero-crossing smoothing outlier method is used to achieve data preprocessing; For small sample data, the Gaussian distribution anomaly recognition model is used to identify abnormal data. First, the Gaussian distribution key parameters are learned through the starfinch optimization algorithm to establish the Gaussian distribution anomaly recognition model. Then, the k-sigma method is used to learn the normal probability threshold of the current data. The IQR method is further used to determine the proportion threshold of the normal current data segment, and finally the abnormal data is screened out. For large sample data, the crow search optimization stacked autoencoder reconstruction method is used to realize abnormal data identification. First, the stacked autoencoder is trained with normal current data, and then the crow search algorithm is used to optimize the parameters of the stacked autoencoder. Finally, the trained stacked autoencoder is applied to the data to be detected to identify abnormal data. For abnormal data, a collaborative multi-source domain adaptive method is used to extract the invariant features of the fault sample domain to achieve fault classification. First, the abnormal samples of the source domain are used to train an independent source domain model and upload it to the target domain. Then, the loss function is used in the abnormal samples of the target domain to increase the recognition rate of the independent target domain model. The independent source domain model and the independent target domain model are aggregated to obtain the overall model, which is finally applied to the abnormal data in the target domain to achieve fault classification. For normal data, piecewise linear fitting is performed, and the time when sub-healthy equipment will fail is predicted based on the parameters of the spliced ​​trend fitting line. First, the current characteristic data is segmented and linear fitting is performed to obtain the changing trend line of the current characteristics. Then, the sparrow search algorithm is used to optimize the parameters of the trend line. The optimized trend lines are spliced ​​to obtain the trend fitting line of the spliced ​​data. Finally, the parameters of the trend line and the warning threshold of the current characteristic amplitude are combined to predict the time when sub-healthy equipment will fail.

2. The motor composite fault identification method as claimed in claim 1, characterized in that: The zero-crossing smoothing outlier method is used to implement data preprocessing, including: The collected high-frequency current data is subjected to zero-crossing judgment, outlier detection and data smoothing to remove abnormal and redundant data and reduce data noise. After preprocessing, a sample data set for motor composite fault diagnosis is formed.

3. The motor composite fault identification method as claimed in claim 1, characterized in that: For small sample data, the Gaussian distribution anomaly recognition model is used to identify abnormal data, including: For small sample data, the starfinch optimization algorithm is used to learn the key parameters of the Gaussian distribution from the data, and a Gaussian distribution anomaly recognition model is established. The k-sigma method is used to learn the normal probability threshold, and the probability value of each data point under the Gaussian distribution is calculated. The proportion of the probability value of each data point in the statistical segment that is less than the normal probability threshold is calculated, and the IQR method is used to determine the proportion threshold of the normal data segment, and abnormal data is identified based on the proportion threshold.

4. The motor composite fault identification method as claimed in claim 1, characterized in that: For large sample data, the crow search optimization stacked autoencoder reconstruction method is used to realize abnormal data identification, including: For large sample data, normal data is used to train the Crow Search Optimized Stacked Autoencoder. The Crow Search Optimized Stacked Autoencoder learns the reconstruction error distribution of normal data and obtains the reconstruction error threshold of normal data. The data to be detected is input into the trained autoencoder, the reconstruction error of the data to be detected is calculated, and abnormal data is identified based on the reconstruction error threshold.

5. The motor composite fault identification method as claimed in claim 1, characterized in that: For abnormal data, a collaborative multi-source domain adaptive method is used to extract the invariant features of the fault sample domain to achieve fault classification, including: Collect samples from abnormal data source domain and target domain, adopt a collaborative multi-source domain adaptive method to extract invariant features of fault sample domain, use the extracted invariant features of fault sample domain to train the classifier, and apply the trained classifier to the target domain data to realize fault classification of samples.

6. The motor composite fault identification method as claimed in claim 1, characterized in that: For normal data, piecewise linear fitting is performed, and the time when sub-healthy equipment will fail is predicted based on the spliced ​​trend fitting line parameters, including: Current characteristics are extracted from high-frequency current data, and the characteristic data are divided into multiple segments. A linear function is used to fit the current characteristic values ​​of each segment to obtain the changing trend line of the current characteristics. The sparrow search algorithm is used to optimize the slope and intercept of the trend line. The trend lines of each segment are spliced ​​in sequence from beginning to end to form a trend fitting line of data splicing. The quantum genetic algorithm is used to obtain the warning threshold of the current characteristic amplitude. The time when sub-healthy equipment will fail is predicted based on the three parameters of the slope, intercept and warning threshold of the current characteristic amplitude of the trend fitting line.

Images