A motor fault diagnosis method based on uniformity and kurtosis calculation
Through uniformity and kurtosis calculation and AGA-CS hybrid classifier model, the problem of high complexity and focus on single faults of existing motor fault detection methods is solved, and the low-cost and fast online failure of induction motors is achieved. Testing improves the accuracy and adaptability of diagnosis.
Patent Information
- Application Number
- CN202211597697.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-13
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-12-13
AI Technical Summary
Existing motor fault detection methods have high mathematical complexity and large computing costs, and most methods focus on detecting a single specific fault, making it difficult to achieve online detection of early multiple faults.
Using a motor fault diagnosis method based on uniformity and kurtosis calculation, the AGA-CS hybrid classifier model is constructed, and the adaptive genetic algorithm and cuckoo search algorithm are optimized. Neural network weights enable diagnosis of various induction motor failures.
It reduces the computational complexity and processing time, realizes effective online detection of various induction motor faults, and improves the accuracy and adaptability of diagnosis.
Smart Images

Figure CN115856623B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of motor health status diagnosis, and in particular to a motor fault diagnosis method based on uniformity and kurtosis calculation. Background Art
[0002] Motors are the foundation of the electric power industry and manufacturing industry, and induction motors have always been important components of various electrical equipment and manufacturing processes. Their main features are low cost, good rigidity and quality performance. However, the economic and technical consequences of unexpected downtime caused by induction motor failures are relatively serious. Designing a reliable early motor fault diagnosis method is a low-cost task, because early detection of motor faults can prevent catastrophic damage to the machine.
[0003] In the past few years, research on induction motor fault detection has aroused great interest. In order to prevent catastrophic failures of motors, many scholars have done a lot of research work. However, many existing methods are based on the analysis of the transient starting current of the motor. The disadvantage of such methods is that their development process has high mathematical complexity and large computational cost. In addition, most existing fault detection methods only focus on detecting a single specific fault, and are usually based on monitoring and analyzing current and vibration signals. These detection methods and techniques are generally suitable for diagnosing independent induction motor faults. In the implementation process, most fault diagnosis methods also rely on complex mathematical algorithms. For example, some frequency domain processing methods need to bring the time domain signal into the frequency domain, and then return to the time domain after processing, which requires a long execution time and computing resources for signal processing, which increases the difficulty of implementation. Summary of the invention
[0004] Purpose of the invention: In view of the high complexity and high computational cost of traditional fault detection methods, as well as the fact that common methods are too focused on detecting a single specific fault, the present invention proposes a motor fault diagnosis method based on uniformity and kurtosis calculation, which has lower computational complexity, shorter processing time, and is suitable for online detection.
[0005] Technical solution: A motor fault diagnosis method based on uniformity and kurtosis calculation includes the following steps:
[0006] Step 1: Calculate the uniformity and kurtosis of the motor transient starting current signal to obtain the motor fault characteristic parameters;
[0007] Step 2: construct a hybrid classifier model based on AGA-CS, train the hybrid classifier model using uniformity and kurtosis characteristic parameters, predict errors through initialized weights, select parent weights according to the predicted errors, complete two algorithm processing on the selected parent weights, one is an adaptive genetic algorithm AGA, and the other is a cuckoo search algorithm CS, and finally merge the results of the two methods to obtain optimized neural network weights and the best solution, thereby constructing a motor fault diagnosis model based on uniformity, kurtosis calculation and AGA-CS;
[0008] Step 3: Implement fault diagnosis of induction motor based on uniformity, kurtosis calculation and AGA-CS motor fault diagnosis model.
[0009] Furthermore, the motor fault characteristic parameters in step 1 include the following:
[0010] Set the fault types to four, namely, motor fault 1BRB with one broken rotor bar, motor fault 2BRB with two broken rotor bars, motor fault BRB with damaged bearing outer ring, and motor fault UNB with unbalanced mechanical load;
[0011] (1) Broken rotor bar fault BRB
[0012] Using the relevant frequency parameter f BRB Indicates the existence of BRB fault, as shown in the following formula:
[0013] f BRB =(1-2kb)·f s , k = 1, 2, 3, ...
[0014] Where k is an integer, f s is the main frequency component of the power supply, b represents the motor slip, and its value range is 0 to 1;
[0015] (2) Bearing failure BRN
[0016] The outer raceway related frequency f0 is determined by the following formula:
[0017]
[0018] Where n is the number of rolling elements, f r is the rotation frequency of the shaft, BD is the ball diameter, PD is the bearing race diameter, and β is the angle between the rolling elements in the race;
[0019] (3) Mechanical imbalance fault UNB
[0020] The rotor is in an unbalanced state, generating an unbalanced force u, which is given by the following formula:
[0021] u=n×r
[0022] Here, m is the mass and r is the eccentricity.
[0023] When the unbalanced force fluctuates with the speed and drags the rotor from the center of the stator to other positions, the mutual inductance between the stator and rotor circuits becomes uneven due to the rotor imbalance, resulting in an induced frequency component f in the stator current. unb , as shown in the following formula:
[0024] f unb =f s [1±k(1-b) / p],k=1,2,3,…
[0025] Among them, f s is the fundamental frequency of the current source, k is an integer, b is the induction motor slip, and p is the number of pole pairs.
[0026] Furthermore, in step 1, the uniformity H and kurtosis K of the motor current signal are extracted. urt Features:
[0027] The uniformity and kurtosis parameters under each motor state are calculated and normalized; statistical analysis is performed on each state to obtain the average μ and standard deviation σ of the uniformity parameters and kurtosis parameters of the current signals of the healthy motor, the motor with one broken rotor bar fault 1BRB, the motor with two broken rotor bar faults 2BRB, the motor with bearing outer ring damage fault BRB, and the motor with unbalanced mechanical load fault UNB; the calculation steps are as follows:
[0028] (1) Calculation of uniformity characteristic parameters
[0029] The uniformity parameter H is described as follows:
[0030]
[0031] Where p(i, j) is the element with position number (i, j) in the normalized GLCM. The uniformity parameter can be used as an index for motor fault detection and classification. The reason is that in the fault current signal of the induction motor, there are different frequency characteristic signals related to the fault, which will change its uniformity, so the uniformity parameter will also change accordingly;
[0032] (2) Calculation of kurtosis characteristic parameters
[0033] The kurtosis characteristic parameter can measure the deviation of the probability distribution and is the fourth-order moment that describes the shape of the signal probability distribution. The kurtosis parameter is calculated as follows:
[0034]
[0035] Where N is the number of samples, x iis the original signal sample, i = 1, 2, 3, ..., N, μ is the random event X = [x1, x2, x3, ..., x N ] is the mean value and σ is the standard deviation.
[0036] Furthermore, the specific steps of weight initialization, error prediction, and selecting parent weights by prediction error in step 2 are as follows:
[0037] Initialize the neural network weights in a random manner, and then use the initialized weights to predict the error, as shown in the following formula:
[0038] E=A V -H
[0039] Among them, E is the error, H is the activation function, and A V is the actual value;
[0040] The activation function H can be obtained by the following formula:
[0041]
[0042] Among them, M is the number of neurons in the hidden layer, N is the number of neurons in the input layer, is the jth weight layer between hidden and output, is the weight between the jth input layer and the jth hidden layer neuron, x i is the i-th input value;
[0043] Finally, the parent weights are selected based on the optimization purpose of reducing the error, and the parent weights are selected through error prediction to further reduce the calculation error.
[0044] Furthermore, the operation of the adaptive genetic algorithm AGA on the selected parent weight in step 2 is specifically as follows:
[0045] According to the proximity of individual fitness value to the group average fitness value and maximum fitness value, the crossover rate and mutation rate are linearly adjusted as follows:
[0046] (1) Completion crossover rate m a The improvement of the crossover rate is:
[0047]
[0048] In the formula, g b is the larger fitness value of the two crossover individuals, g max is the maximum individual fitness value in the population, g avg is the average fitness value of all individuals in the population, a1 and a2 are random numbers between 0 and 1;
[0049] (2) Complete mutation rate m rImprovement and adaptive mutation, the mutation rate m r Improved to:
[0050]
[0051] In the formula, g is the fitness value of the mutant individual, g max is the maximum individual fitness value in the population, g avg is the average fitness value of all individuals in the population, a4 is a random number between 0 and 1, and a3 is set as follows:
[0052]
[0053] Where R is a random value between [0, 1] that changes in each iteration, and c and μ are mutation parameters. After executing the mutation process, we will obtain a complete new offspring. Next, the AGA algorithm obtains new optimization weights through the new offspring, and repeats (1) and (2) to obtain a set of new optimization weights based on AGA.
[0054] Furthermore, the operation of the cuckoo algorithm CS for the selected parent weight in step 2 is specifically as follows:
[0055] During execution, the locations of the cuckoo nests are mapped to solutions in the algorithm population space. The specific steps are as follows:
[0056] (1) Initialization: randomly generate N solutions to the problem and record the optimal solution;
[0057] Let N e = {N e1 , N e1 , N e1 ,…,N em} is a candidate solution, m is the dimension of the optimization problem, the algorithm uses a random walk to complete the search, and the quality of the bird's nest location is used as the fitness, which is calculated as follows:
[0058]
[0059] in, represents the i-th bird's nest to be updated in the t-th generation, is the updated next generation of bird nests, ρ is the step size scaling factor, is a dot product operation, L(λ) is the Levy random flight path, representing the jump path of Levy flight, whose direction and length are uncertain; the relationship between the continuous Levy flight jump path and time obeys the Levy distribution. By simplifying the distribution function and Fourier transforming, the corresponding probability density function can be obtained, as shown in the following formula:
[0060] L(λ)~u=t -λ ; 1<λ<3
[0061] Among them, λ represents the power coefficient;
[0062] (2) Step length evaluation to obtain the flight jump path:
[0063] Evaluate the obtained step size of each parent weight and calculate the step size using the following expression:
[0064]
[0065] Among them, S Z is the step size, α is the step size parameter, w t is the current parent weight, is the best solution so far. r is a random number between [0, 1] from a standard normal distribution. S is the number of steps. The S value is obtained by simulating the Levy flight jump path using the following formula:
[0066]
[0067] Among them, S represents the Levy flight jump path, which is equivalent to L(λ), β is a parameter between [1, 2], u and v are normally distributed random numbers, and their distribution is shown in the following formula
[0068]
[0069] The standard deviation of the corresponding normal distribution is σ u With σ v , the specific definition is as follows:
[0070]
[0071] (3) Use Levy flight to update the position of the solution, obtain the new candidate solution position, and then use the greedy selection strategy to retain the better solution;
[0072] (4) According to the discovery probability P a , discard some solutions and generate new solutions through random walk to replace the discarded solutions;
[0073] According to the step size S of each parent weight Z And search path, the cuckoo individual completes the flight process. Since its search path changes randomly, the cuckoo will switch from one area to another, which enhances its overall optimization ability. After completing the flight walk of global exploration, based on the current solution, the preferred random walk method is used to complete the local search. Compare the discovery probability P a With threshold constant like It means that the bird's nest is found, that is, the current solution is a poor solution, and then the bird's nest parameters are updated, the poor solutions are eliminated, and new candidate solutions are generated, thereby leaving the best cuckoo individual; the bird's nest position update formula is described as follows:
[0074]
[0075] in, Indicates that the bird's nest is waiting to be updated. is the updated bird's nest, ξ is a random number that satisfies uniform distribution between [0, 1), are any two bird nests of generation t;
[0076] (5) If the termination condition is met, output the best solution; otherwise, return to (3) to continue iterating;
[0077] (6) Generate a new solution. Create a new optimization solution and update the parent weight based on the CS algorithm. The specific formula is as follows:
[0078] w t+1 =w t +S z
[0079] Among them, w t+1 is the new weight, S z is the step length, w t is the current parent weight.
[0080] Furthermore, the specific steps of implementing the fault diagnosis of the induction motor based on the uniformity, kurtosis calculation and the motor fault diagnosis model of the AGA-CS hybrid classifier in step 3 are:
[0081] The transient starting current signal of the motor is obtained, and then it is processed and analog-to-digital converted. The desired features, including uniformity parameters and kurtosis parameters, are obtained from the obtained discrete current signal, and are used as the input of a hybrid classifier model based on AGA-CS. In the hybrid classifier based on AGA-CS, the adaptive genetic algorithm AGA and the cuckoo search algorithm CS are mixed and the neural network weights are used as input, the initialization weights are evaluated as fitness, and the parent weights are selected according to the fitness values. Then, the selected parent weights are processed by the two algorithms, and finally the results of the two methods are merged to obtain the optimized neural network weights and the best solution, which are used to diagnose the faults of induction motors, and finally the correct classification of five motor states is obtained, including five states of a healthy motor, a motor with one broken rotor bar fault 1BRB, a motor with two broken rotor bar faults 2BRB, a motor with a bearing outer ring damage fault BRB, and a motor with an unbalanced mechanical load fault UNB.
[0082] Beneficial effects:
[0083] The present invention completes the extraction of motor fault characteristic parameters based on uniformity and kurtosis parameters, and completes feature extraction only by detecting the uniformity and kurtosis of the single-phase transient starting current of the motor, thereby effectively reducing the computational complexity of the algorithm, shortening the processing time, and being more suitable for online processing, which is a low-cost computing technology; on the other hand, a hybrid classifier model based on AGA-CS is established, and the error fitness value of the neural network is used to select the parent weight, and the selected parent weight is processed by two algorithms, one is an adaptive genetic algorithm (AGA), and the other is a cuckoo search algorithm (CS), and finally the results of the two methods are merged to obtain the optimized neural network weight and the best solution, and the classification ability of the algorithm is further improved; finally, a motor fault diagnosis model based on uniformity, kurtosis and AGA-CS is constructed, and effective prediction of five faults of the induction motor is realized, and the diagnosis objects are healthy motors, motor faults with one broken rotor bar (1BRB), motor faults with two broken rotor bars (2BRB), motor faults with damaged bearing outer rings (BRB) and motor faults with unbalanced mechanical loads (UNB).
[0084] Since the present invention focuses on the frequently occurring motor fault problems and achieves the purpose of detecting multiple specific faults at the same time, it is more adaptable. Future work will focus on evaluating the multi-induction motor fault detection technology in different scenarios, considering combining other fault conditions and signal inspection techniques to recover more signal features, thereby completing higher-precision fault identification and classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0085] Figure 1 is the probability density function (PDF) of different motor motion states based on kurtosis;
[0086] Figure 2 is the probability density function (PDF) of different motion states of the motor based on uniformity;
[0087] Figure 3 It is a conventional neural network ANN structure;
[0088] Figure 4 It is the execution process of the AGA-CS hybrid classifier model;
[0089] Figure 5 It is the implementation process of the motor fault diagnosis model based on uniformity, kurtosis and AGA-CS;
[0090] Figure 6 It is the average effect of different classification algorithms on motor fault diagnosis. DETAILED DESCRIPTION
[0091] In order to better explain the present invention and facilitate understanding, the technical solution of the present invention is described in detail below. The following examples are explanations of the present invention, but the present invention is not limited to the following examples.
[0092] The present invention provides a motor fault diagnosis method based on uniformity and kurtosis calculation. The specific implementation process includes the following steps:
[0093] Step 1: Complete the extraction of motor fault parameters based on uniformity and kurtosis parameters. Based on the analysis of the uniformity and kurtosis of the motor transient starting current signal, a simpler characteristic parameter that can effectively characterize the motor fault is obtained.
[0094] Step (1) a: Select the induction motor fault diagnosis object and type, collect three-phase current signals, and obtain motor fault samples;
[0095] Five motor states are set, namely healthy motor, motor fault 1BRB with one broken rotor bar, motor fault 2BRB with two broken rotor bars, motor fault BRB with bearing outer ring damage, and motor fault UNB with unbalanced mechanical load. The fault type is determined based on the following:
[0096] (1) Broken rotor bar fault BRB
[0097] BRB is the most common type of motor fault, usually caused by overload, thermal imbalance, electromagnetic force, noise, vibration, environmental damage or manufacturing process, etc. This fault is difficult to diagnose. When a broken rotor bar fault BRB occurs, it will cause rotor asymmetry, causing the motor to operate under abnormal operating conditions, resulting in the induction motor generating a reverse sequence of rotor currents, which in turn introduces a unique component in the stator current spectrum. At this time, the relevant frequency parameter f BRB To indicate the existence of BRB fault, as shown in formula (1):
[0098] f BRB =(1-2kb)·f s , k = 1, 2, 3, ... (1)
[0099] Where k is an integer, f s is the main frequency component of the power supply, b represents the motor slip, and its value range is 0 to 1.
[0100] (2) Bearing failure BRN
[0101] When any component of the motor bearing fails, such as the inner raceway, outer raceway or rolling element of the bearing, it will cause specific components in the vibration or current signal of the induction motor. These components are related to both the power supply frequency and the mechanical system frequency. In addition, the characteristic frequency associated with the bearing fault depends on the bearing geometry and the machine speed. For example, when there is a defect in the outer ring, the outer ring defect will cause a specific pulse every time the rolling element contacts the defect. In theory, the outer raceway related frequency f0 can be determined by formula (2):
[0102]
[0103] Where n is the number of balls (rolling elements), f r is the rotation frequency of the shaft, BD is the ball diameter, PD is the bearing race diameter, and β is the angle between the rolling elements in the race.
[0104] (3) Mechanical imbalance fault UNB
[0105] When the mechanical load of an induction motor is unevenly distributed, if the center of mass is placed outside the motor's rotating shaft, the motor will have an unbalanced fault. Defects in the manufacturing process are the main cause of rotor unbalance failures. In addition, thermal expansion can also cause internal misalignment or shaft deviation that affects the rotor, resulting in rotor unbalance failures. When the weight distribution near the rotor's rotating center is uneven, an unbalanced state will occur, generating an unbalanced force u, which is specifically given by equation (3):
[0106] u=m×r (3)
[0107] Here, m is the mass and r is the eccentricity.
[0108] When the unbalanced force fluctuates with the speed and drags the rotor from the center of the stator to other positions, the mutual inductance between the stator and rotor circuits becomes uneven due to the rotor imbalance, resulting in an induced frequency component f in the stator current. unb , as shown in formula (4):
[0109] f unb =f s [1±k(1-b) / p], k=1, 2, 3,… (4)
[0110] Among them, f s is the fundamental frequency of the current source, k is an integer, b is the induction motor slip, and p is the number of pole pairs.
[0111] Step (1)b: Extract the uniformity H and kurtosis K of the motor state current signal urt Characteristic parameters.
[0112] Calculate the uniformity parameter H and kurtosis parameter K for each motor state urt, and normalize it, and then perform statistical analysis on each state to obtain the uniformity parameter H and kurtosis parameter K of the current signal of the healthy motor, the motor with one broken rotor bar fault 1BRB, the motor with two broken rotor bar faults 2BRB, the motor with bearing outer ring damage fault BRB, and the motor with unbalanced mechanical load fault UNB urt The mean μ and standard deviation σ.
[0113] The steps for calculating the uniformity and kurtosis characteristic parameters are as follows:
[0114] (1) Calculate the uniformity characteristic parameter H
[0115] The uniformity parameter H is a texture property that estimates the grayscale variation of image pixels. It is derived from the gray-level co-occurrence matrix (GLCM) of the image and is used to measure the compactness of the elements distributed on the diagonal in the GLCM. Through the grayscale function of any other pixel, the GLCM can show the number of occurrences of the grayscale of each pixel at a predetermined geometric position. Its uniformity range is 0-1. When the diagonal element value is 1, the uniformity reaches the maximum value. The following formula can be used to describe the uniformity parameter H:
[0116]
[0117] Where p(i, j) is the element with position number (i, j) in the normalized GLCM. The uniformity parameter can be used as an index for motor fault detection and classification because there are different frequency characteristic signals related to the fault in the fault current signal of the induction motor, which will change its uniformity, so the uniformity parameter will also change accordingly.
[0118] (2) Calculate the kurtosis characteristic parameter K urt
[0119] Kurtosis parameter K urt It can measure the deviation of the probability distribution, which is the fourth-order moment that describes the shape of the signal probability distribution. It can also measure the slight differences between distributions with different shapes. Therefore, it can be used as an indicator for motor fault diagnosis. If the kurtosis parameter has a high pulse fault component, the signal will have a sharp signal strength distribution, so there will be a high kurtosis value. The kurtosis parameter of a random event x is calculated as follows:
[0120]
[0121] Where N is the number of samples, x i is the original signal sample, i = 1, 2, 3, ..., N, μ is the random event X = [x1, x2, x3, ..., x N ] is the mean value and σ is the standard deviation.
[0122] In the specific implementation, the uniformity parameter H and the kurtosis parameter K under each motion condition are obtained by equations (5) and (6) respectively. urt , and normalize it. Perform statistical analysis on each motion state. Obtain the uniformity parameter H and kurtosis parameter K of the current signal of the healthy motor, the motor with one broken rotor bar fault 1BRB, the motor with two broken rotor bar faults 2BRB, the motor with bearing outer ring damage fault BRB, and the motor with unbalanced mechanical load fault UNB urt The mean μ and standard deviation σ of the signal are used to indicate the probability density function PDF of the corresponding signal.
[0123] Figure 1 , Figure 2 It is the probability density function PDF based on uniformity and kurtosis characteristics. It can be seen that there is an obvious overlap between the processing conditions. Finally, the classifier is used to complete the classification operation of the motor fault to accurately identify the fault type of the motor.
[0124] Step 2: Complete the construction of the hybrid classifier model based on AGA-CS. Predict the error by initializing the weights, select the parent weights according to the prediction error, and complete the processing of two algorithms on the selected parent weights, one is the adaptive genetic algorithm AGA, and the other is the cuckoo search algorithm CS. Finally, the results of the two methods are combined to obtain the optimized neural network weights and the best solution.
[0125] Artificial neural network ANN replicates neural structure and human brain function, has strong flexibility and adaptability, and generally consists of a single input layer and a single output layer. Figure 3 .
[0126] In a neural network, the training process is very important, and the training algorithm used can obtain the best weight. The common training algorithm is the back propagation algorithm, which has a slow convergence speed, a large amount of calculation, and often falls into a local optimum. Therefore, the present invention develops a new neural network training method, namely a hybrid classification method AGA-CS based on an adaptive genetic algorithm AGA and a cuckoo search algorithm CS. The adaptive genetic algorithm AGA and the cuckoo search algorithm CS are mixed and used, and the weight is used as input, the initialization weight is evaluated as fitness, and the parent weight is selected according to the fitness value. Then, the selected parent weight is processed by two algorithms, one is the adaptive genetic algorithm AGA, and the other is the cuckoo search algorithm CS. Finally, the results of the two methods are merged to obtain the optimized neural network weight and the best solution.
[0127] Figure 4 It is the execution process of the AGA-CS hybrid classifier model, combined with Figure 4 , the main implementation process of step 2 is as follows:
[0128] Step (2)a: weight initialization, error prediction, and selecting parent weights by predicting the error;
[0129] Initialize the neural network weights in a random manner, and then predict the error by initializing the weights, as shown in formula (7):
[0130] E=A V -H (7)
[0131] Among them, E is the error, H is the activation function, and A V is the actual value.
[0132] The activation function H can be obtained by using formula (8):
[0133]
[0134] Among them, M is the number of neurons in the hidden layer, N is the number of neurons in the input layer, is the jth weight layer between hidden and output, is the weight between the neurons in the ith input layer and the jth hidden layer, x i is the ith input value.
[0135] Finally, the parent weights are selected based on the optimization purpose of reducing the error, and the parent weights are selected through error prediction to further reduce the calculation error.
[0136] Step (2) b: completing the operation of the adaptive genetic algorithm AGA;
[0137] The crossover rate and mutation rate of the genetic algorithm GA directly affect the convergence speed of the algorithm. On the one hand, if the crossover rate is too high, it is difficult to iterate the genes of excellent individuals, and it will cause the speed of new individuals to slow down, both of which will cause the algorithm to stop and converge prematurely; on the other hand, the greater the mutation rate, the algorithm will become a state of complete random search, and whether the global optimal result can be searched will no longer be affected by operations such as selection, crossover and mutation. On the contrary, the smaller the mutation rate, the less likely it is to produce new excellent individuals. Therefore, when the standard GA is implemented, it is necessary to use repeated experiments to find targeted crossover rates and mutation rates for each optimization problem, but this is almost impossible to complete, and it will eventually lead to difficulty in finding the optimal parameters. Therefore, the present invention adopts an adaptive genetic algorithm AGA, and according to the proximity of the individual fitness value to the group average fitness value and the maximum fitness value, the crossover rate and mutation rate are linearly adjusted to improve the convergence speed of the algorithm. It is as follows:
[0138] (1) Completion crossover rate m a Improvements
[0139] The crossover rate is for the new offspring of the genetic algorithm, which is formed by combining the chromosomes of the parents. On the one hand, if the larger fitness of the two individuals to be crossed is significantly different from the maximum fitness of the population, the crossover individual is regarded as a non-high-quality individual, and its crossover rate needs to be appropriately increased to increase the probability of obtaining high-quality samples through crossover; on the other hand, if the difference between the best fitness of the population and the average fitness is large, the sample is considered to be relatively scattered, which means that the fitness function has not yet reached the extreme value, and its crossover rate needs to be appropriately reduced to make the scattered population iterate as soon as possible. In addition, when the better fitness value between the crossover individuals is less than the average fitness of the population, the sample is regarded as a poor sample. In order to reduce the amount of calculation, the crossover rate of the poor sample will no longer be considered based on the population distribution and will be set to a constant value. According to the above principles, the crossover rate is improved to:
[0140]
[0141] In the formula, g b is the larger fitness value of the two crossover individuals, g max is the maximum individual fitness value in the population, g avg is the average fitness value of all individuals in the population, a1 and a2 are random numbers between 0 and 1, usually set manually.
[0142] (2) Complete mutation rate m r Improvement and adaptive variation
[0143] The purpose of adaptive mutation is to obtain complete new offspring. In genetic algorithms, the adaptive mutation rate is randomly selected. Similar to the crossover rate, it is necessary to consider adjusting the difference between the mutant individual and the average fitness, so the mutation rate mr is improved to:
[0144]
[0145] In the formula, g is the fitness value of the mutant individual, g max is the maximum individual fitness value in the population, g avg is the average fitness value of all individuals in the population, a4 is a random number between 0 and 1, and a3 is set as follows:
[0146]
[0147] Where R is a random value between [0, 1] that changes in each iteration, and c and μ are mutation parameters (in the implementation, c=μ=2 is set). After the mutation is performed, a complete new offspring will be obtained. Next, the AGA algorithm obtains new optimization weights through the new offspring. Repeating (1) and (2) will obtain a set of new optimization weights based on AGA.
[0148] Step (2) c: complete the operation of the cuckoo algorithm CS;
[0149] Like all evolutionary algorithms, the cuckoo optimization algorithm CS uses a randomly generated population for iterative search. During execution, the locations of the cuckoo's parasitic nests are mapped to solutions in the algorithm population space. The specific steps are as follows:
[0150] (1) Initialization: randomly generate N solutions to the problem and record the optimal solution.
[0151] Let N e = {N e1 , N e1 , N e1 ,…,N em} is a candidate solution, m is the dimension of the optimization problem, the algorithm uses a random walk to complete the search, and the quality of the bird's nest location is used as the fitness. The specific calculation is as shown in formula (12):
[0152]
[0153] in, represents the i-th bird's nest to be updated in the t-th generation, is the updated next generation bird nest, ρ is the step size scaling factor, and in implementation, its value is set to ρ=0.01, is the dot product operation, L(λ) is the Levy random flight path, representing the jump path of Levy flight, whose direction and length are uncertain. The relationship between the continuous Levy flight jump path and time obeys the Levy distribution. By simplifying the distribution function and Fourier transforming, the corresponding probability density function can be obtained, as shown in formula (13):
[0154] L(λ)~u=t -λ ; 1<λ<3 (13)
[0155] Here, λ represents the power coefficient.
[0156] (2) Step length evaluation to obtain the flying jump path.
[0157] In this step, the weight steps of each parent obtained in step 3 are evaluated and the step length is calculated using the following expression:
[0158]
[0159] Among them, S Z is the step size, α is the step size parameter (in the implementation, α is set to 0.02), w t is the current parent weight, is the best solution so far. r is a random number between [0, 1] from a standard normal distribution. S is the step number. By simulating the Levy flight jump path, the s value is obtained using formula (15):
[0160]
[0161] Where S represents the Levy flight jump path, equivalent to L(λ), β is a parameter between [1, 2]. In the implementation, β=1.6 is set, u and v are normally distributed random numbers, and their distribution is shown in formula (16). The standard deviation of the corresponding normal distribution is σ u With σ v , the specific definition is shown in formula (17):
[0162]
[0163]
[0164] (3) Use Levy flight to update the position of the solution, obtain the new candidate solution position, and then use the greedy selection strategy to retain the better solution.
[0165] (4) According to the discovery probability P a , discard some solutions and generate new solutions through random walk to replace the discarded solutions.
[0166] According to the step size S of each parent weight Z The cuckoo individual completes the flight process. Since its search path changes randomly, the cuckoo will switch from one area to another, which enhances its overall optimization ability. After completing the global exploration flight walk, based on the current solution, the preferred random walk method is used to complete the local search. Compare the discovery probability P a With threshold constant like (In implementation, set ), it means that the bird's nest is found, that is, the current solution is a poor solution, then the bird's nest parameters are updated, the poor solution is eliminated, and a new candidate solution is generated, thus leaving the best cuckoo individual. The bird's nest position update formula is as follows (18):
[0167]
[0168] in, Indicates that the bird's nest is waiting to be updated. is the updated bird's nest, ξ is a random number that satisfies uniform distribution between [0, 1), are any two bird nests of generation t.
[0169] (5) If the termination condition is met, output the best solution; otherwise, return to (3) to continue iterating.
[0170] (6) Generate new solutions.
[0171] In this step, a new optimization scheme will be created to complete the update of the parent weight based on the CS algorithm, as follows:
[0172] w t+1 =w t +S z (19)
[0173] Among them, w t+1 is the new weight, S z is the step length, w t is the current parent weight.
[0174] Step (2)d: hybrid algorithm merging;
[0175] The algorithm results of step (2)b and step (2)c are combined, and the combined weights are regarded as optimized weights to obtain the best solution.
[0176] Step (2) e: determine the termination criteria;
[0177] The maximum number of weights for training the neural network is considered as the termination criterion to determine whether the termination condition is reached. If not, go to the error prediction in step 1 and continue to perform the optimization until the termination condition is reached to obtain the optimal solution.
[0178] Step 3: Construct a motor fault diagnosis model based on uniformity, kurtosis and AGA-CS, and achieve correct prediction of five induction motor faults.
[0179] Figure 5 This is the implementation process of the motor fault diagnosis model based on uniformity, kurtosis and AGA-CS. The transient starting current signal of the motor is obtained by a current clamp, and then it is converted into an analog-to-digital signal. The uniformity parameter and kurtosis parameter are obtained from the obtained discrete current signal and used as the input of the AGA-CS hybrid classifier. The adaptive genetic algorithm AGA and the cuckoo search algorithm CS are mixed and the neural network weights are used as input. The initialization weights are evaluated as fitness, and the parent weights are selected according to the fitness value. Then, the selected parent weights are processed by the two algorithms. Finally, the results of the two methods are merged to obtain the optimized neural network weights and the best solution, and finally the correct classification of five motor states is achieved, including five states of a healthy motor, a motor with a broken rotor bar fault 1BRB, a motor with two broken rotor bars faults 2BRB, a motor with a bearing outer ring damage fault BRB, and a motor with an unbalanced mechanical load fault UNB.
[0180] Step 4: Based on the above technology, the MATLAB neural network toolbox is selected as the implementation tool, an experimental platform is built, a motor fault diagnosis model based on uniformity, kurtosis and AGA-CS is established, and the classification of five motor states is completed. The following test experiments are mainly completed:
[0181] (1) Motor fault diagnosis experiment based on uniformity, kurtosis and AGA-CS
[0182] The transient starting current signal of the motor is used to identify and classify the state of the motor. There are five diagnostic objects, namely healthy motor, motor fault 1BRB with one broken rotor bar, motor fault 2BRB with two broken rotor bars, motor fault BRB with damaged outer bearing ring, and motor fault UNB with unbalanced mechanical load. The experimental object is an induction motor with model WEG 00136APE48T. Its working power supply is 220V AC, frequency is 60Hz, and it has 28 rotors. The mechanical load used is an ordinary AC generator, which is equivalent to one-fourth of the nominal load of the induction motor under test. In the specific implementation, the three-phase current signal of the induction motor is collected by an AC current clamp, and the collected data is converted into A / D using a 16-bit analog-to-digital converter (ADS7809). During the implementation, it is ensured that one motor experimental object is in a healthy state, and the 1BRB / 2BRB with one and two broken rotor bars is generated by drilling one hole and two holes on the rotor shaft of the experimental motor object, respectively, with a hole diameter of 8mm. In addition, a 2mm hole is drilled on the motor bearing by using a drill bit to damage it artificially, so that the outer ring of the bearing is damaged, thereby generating the motor fault BRB with damaged outer ring of the bearing. Finally, the mechanical unbalance fault UNB is generated by adding mass to the pulley arm. In the experiment, 50 experiments were conducted for each of the above states, and the sampling frequency of each experiment was 1.5kHz, that is, 4096 samples were obtained at the moment of starting the induction motor. The experimental results are finally obtained as shown in Table 1.
[0183] Table 1 Motor fault detection and classification results based on uniformity and kurtosis features
[0184] Motor status Fault classification accuracy Motor health status 98% A motor fault with a broken rotor bar 1BRB 95% Motor failure with two broken rotor bars 2BRB 96% Motor failure with damaged bearing outer ring BRB 95% Motor Fault with Unbalanced Mechanical Load UNB 94%
[0185] As can be seen from Table 1, the proposed method can effectively identify and classify the state of induction motors. The fault classification accuracy can reach 98%, 95%, 96%, 95% and 94% for motors in a healthy state, motors with one broken rotor bar fault 1BRB, motors with two broken rotor bar faults 2BRB, motors with bearing outer ring damage fault BRB and motors with unbalanced mechanical load fault UNB. Therefore, the proposed technology can reliably classify motor faults. By only analyzing one phase of the motor starting current, the classification effect of the induction motor for different faults can be ensured.
[0186] (2) Comparative experiments on different classifier models
[0187] In the implementation, the hybrid classification method (AGA-CS) proposed in the present invention is experimentally verified with the conventional BP algorithm, as well as the adaptive genetic algorithm (AGA), cuckoo algorithm (CS) and genetic algorithm (GA) used alone. After the model is trained, 30 groups of samples are randomly selected, and the fault samples are sent to the trained model for testing. In the specific implementation process, it is necessary to count the states of five motor states, including healthy motors, motors with one broken rotor bar fault 1BRB, motors with two broken rotor bar faults 2BRB, motors with bearing outer ring damage fault BRB and motors with unbalanced mechanical load fault UNB. Finally, the recognition rates obtained by different methods are averaged, and the diagnosis results are shown in Table 2.
[0188] Table 2 Comparison of diagnostic effects of different classification algorithms for different motor states
[0189]
[0190]
[0191] Figure 6 It is the average effect of different classification algorithms on motor fault diagnosis. As can be seen from the above chart, the prediction result of the motor fault diagnosis method based on uniformity and kurtosis calculation and AGA-CS hybrid classification mode is the best, and its absolute value of prediction error is significantly smaller than the effect of using only one traditional algorithm (AGA, CS, BP, GA). Since the training samples contain motor fault information, and the AGA-CS hybrid classification model can obtain more in-depth fault characterization information, therefore, compared with traditional algorithms, the AGA-CS classifier shows more obvious advantages in motor fault diagnosis, has a higher diagnostic accuracy, and can obtain a diagnostic effect far higher than that of using only one classification algorithm.
[0192] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.
Claims
1. A motor fault diagnosis method based on uniformity and kurtosis calculation, characterized in that: The following steps are involved: Step 1: Calculate the uniformity and kurtosis of the motor transient starting current signal to obtain the motor fault characteristic parameters; Step 2: construct a hybrid classifier model based on AGA-CS, train the hybrid classifier model using uniformity and kurtosis characteristic parameters, predict errors through initialized weights, select parent weights according to the predicted errors, complete two algorithm processing on the selected parent weights, one is an adaptive genetic algorithm AGA, and the other is a cuckoo search algorithm CS, and finally merge the results of the two methods to obtain optimized neural network weights and artificial neural network ANN optimal solutions, thereby constructing a motor fault diagnosis model based on uniformity, kurtosis calculation and AGA-CS; Initialize the neural network weights in a random manner, and then use the initialized weights to predict the error, as shown in the following formula: E=A V -H Among them, E is the error, H is the activation function, and A V is the actual value; The activation function H is obtained by the following formula: Among them, M is the number of neurons in the hidden layer, N is the number of neurons in the input layer, is the jth weight layer between hidden and output, is the weight between the neurons in the ith input layer and the jth hidden layer, x i is the i-th input value; Finally, the parent weight is selected based on the optimization purpose of reducing the error, and the parent weight is selected through error prediction to further reduce the calculation error; The operation of the adaptive genetic algorithm AGA for the selected parent weight is as follows: According to the proximity of individual fitness value to the group average fitness value and maximum fitness value, the crossover rate and mutation rate are linearly adjusted as follows: (1) Completion crossover rate m a The improvement of the crossover rate is: In the formula, g b is the larger fitness value of the two crossover individuals, g max is the maximum individual fitness value in the population, g avg is the average fitness value of all individuals in the population, a1 and a2 are random numbers between 0 and 1; (2) Complete mutation rate m r Improvement and adaptive mutation, the mutation rate m r Improved to: In the formula, g is the fitness value of the mutant individual, g max is the maximum individual fitness value in the population, g avg is the average fitness value of all individuals in the population, a4 is a random number between 0 and 1, and a3 is set as follows: Where R is a random value between [0, 1] that changes in each iteration, and c and μ are mutation parameters. After executing the mutation process, we will obtain a complete new offspring. Next, the AGA algorithm obtains new optimization weights through the new offspring, and repeats (1) and (2) to obtain a set of new optimization weights based on AGA. The operation of the cuckoo algorithm CS for the selected parent weight is as follows: During execution, the locations of the cuckoo nests are mapped to solutions in the algorithm population space. The specific steps are as follows: (1) Initialization: randomly generate N solutions to the problem and record the optimal solution; Let N e = {N e1 ,N e1 ,N e1 ,…,N em } is a candidate solution, m is the dimension of the optimization problem, the algorithm uses a random walk to complete the search, and the quality of the bird's nest location is used as the fitness, which is calculated as follows: in, represents the i-th bird's nest to be updated in the t-th generation, is the updated next generation of bird nests, ρ is the step size scaling factor, is a dot product operation, L(λ) is the Levy random flight path, representing the jump path of Levy flight, whose direction and length are uncertain; the relationship between the continuous Levy flight jump path and time obeys the Levy distribution. By simplifying the distribution function and Fourier transforming, the corresponding probability density function can be obtained, as shown in the following formula: L(λ)~u=t -λ ;1<λ<3 Among them, λ represents the power coefficient; (2) Step length evaluation to obtain the flight jump path: Evaluate the obtained step size of each parent weight and calculate the step size using the following expression: Among them, S Z is the step size, α is the step size parameter, w t is the current parent weight, is the best solution so far. r is a random number between [0,1] from a standard normal distribution. S is the number of steps. The S value is obtained by simulating the Levy flight jump path using the following formula: Among them, S represents the Levy flight jump path, which is equivalent to L(λ), β is a parameter between [1, 2], u and v are normally distributed random numbers, and their distribution is shown in the following formula The standard deviation of the corresponding normal distribution is σ u With σ v , the specific definition is as follows: (3) Use Levy flight to update the position of the solution, obtain the new candidate solution position, and then use the greedy selection strategy to retain the better solution; (4) According to the discovery probability P a , discard some solutions and generate new solutions through random walk to replace the discarded solutions; According to the step size S of each parent weight Z And search path, the cuckoo individual completes the flight process. Since its search path changes randomly, the cuckoo will switch from one area to another, which enhances its overall optimization ability. After completing the flight walk of global exploration, based on the current solution, the preferred random walk method is used to complete the local search. Compare the discovery probability P a With threshold constant like It means that the bird's nest is found, that is, the current solution is a poor solution, and then the bird's nest parameters are updated, the poor solutions are eliminated, and new candidate solutions are generated, thereby leaving the best cuckoo individual; the bird's nest position update formula is described as follows: in, Indicates that the bird's nest is waiting to be updated. is the updated bird's nest, ξ is a random number that satisfies uniform distribution between [0,1), are any two bird nests of generation t; (5) If the termination condition is met, output the best solution; otherwise, return to (3) to continue iterating; (6) Generate a new solution. Create a new optimization solution and update the parent weight based on the CS algorithm. The specific formula is as follows: w t+1 =w t +S z Among them, w t+1 is the new weight, S z is the step length, w t is the current parent weight; Step 3: Implement fault diagnosis of induction motor based on uniformity, kurtosis calculation and AGA-CS motor fault diagnosis model.
2. The motor fault diagnosis method based on uniformity and kurtosis calculation according to claim 1 is characterized in that: In step 1, the motor fault characteristic parameters are set to five motor states, namely, a healthy motor, a motor fault 1BRB with one broken rotor bar, a motor fault 2BRB with two broken rotor bars, a motor fault BRN with a damaged bearing outer ring, and a motor fault UNB with an unbalanced mechanical load; There are four types of faults, and the basis for determining the fault types is: (1) Motor fault 1BRB with one broken rotor bar, motor fault 2BRB with two broken rotor bars Using the relevant frequency parameter f BRB It indicates the existence of motor fault 1BRB with one broken rotor bar and motor fault 2BRB with two broken rotor bars, as shown in the following formula: f BRB =(1-2kb)·f s ,k=1,2,3,… Where k is an integer, f s is the main frequency component of the power supply, b represents the motor slip, and its value range is 0 to 1; (2) Motor fault with bearing outer ring damage BRN The outer raceway related frequency f0 is determined by the following formula: Where n is the number of rolling elements, f r is the rotation frequency of the shaft, BD is the ball diameter, PD is the bearing race diameter, and β is the angle between the rolling elements in the race; (3) Motor fault with unbalanced mechanical load UNB The rotor is in an unbalanced state, generating an unbalanced force u, which is given by the following formula: u=m×r Here, m is the mass and r is the eccentricity. When the unbalanced force fluctuates with the speed and drags the rotor from the center of the stator to other positions, the mutual inductance between the stator and rotor circuits becomes uneven due to the rotor imbalance, resulting in an induced frequency component f in the stator current. unb , as shown in the following formula: f unb =f s [1±k(1-b) / p],k=1,2,3,… Among them, f s is the fundamental frequency of the current source, k is an integer, b is the induction motor slip, and p is the number of pole pairs.
3. The motor fault diagnosis method based on uniformity and kurtosis calculation according to claim 1, characterized in that: In step 1, the uniformity H and kurtosis K of the motor current signal are extracted urt Features: The uniformity and kurtosis parameters were calculated for each motor state and normalized; A statistical analysis is performed on each state to obtain the average μ and standard deviation σ of the uniformity parameter and kurtosis parameter of the current signal of the healthy motor, the motor with a broken rotor bar 1BRB, the motor with two broken rotor bars 2BRB, the motor with a bearing outer ring damaged motor fault BRN, and the motor fault UNB with unbalanced mechanical load; the calculation steps are as follows: (1) Calculation of uniformity characteristic parameters The uniformity parameter H is described as follows: Where p(i,j) is the element with position number (i,j) in the normalized GLCM. The uniformity parameter can be used as an index for motor fault detection and classification. The reason is that in the fault current signal of the induction motor, there are different frequency characteristic signals related to the fault, which will change its uniformity, so the uniformity parameter will also change accordingly; (2) Calculation of kurtosis characteristic parameters The kurtosis characteristic parameter can measure the deviation of the probability distribution and is the fourth-order moment that describes the shape of the signal probability distribution. The kurtosis parameter is calculated as follows: Where N is the number of samples, x i is the original signal sample, i = 1, 2, 3, ..., N, μ is the random event X = [x1, x2, x3, ..., x N ] is the mean value and σ is the standard deviation.
4. A motor fault diagnosis method based on uniformity and kurtosis calculation according to claim 1, characterized in that: The specific steps of implementing the fault diagnosis of the induction motor based on the uniformity, kurtosis calculation and the motor fault diagnosis model of the AGA-CS hybrid classifier in step 3 are: The transient starting current signal of the motor is obtained, and then it is processed and analog-to-digital converted. The desired features, including uniformity parameters and kurtosis parameters, are obtained from the obtained discrete current signal, and are used as the input of a hybrid classifier model based on AGA-CS. In the hybrid classifier based on AGA-CS, the adaptive genetic algorithm AGA and the cuckoo search algorithm CS are mixed and the neural network weights are used as input, the initialization weights are evaluated as fitness, and the parent weights are selected according to the fitness values. Then, the selected parent weights are processed by the two algorithms, and finally the results of the two methods are merged to obtain the optimized neural network weights and the best solution, which are used to diagnose induction motor faults, and finally the correct classification of five motor states is obtained, including five states of a healthy motor, a motor with a broken rotor bar 1BRB, a motor with a broken rotor bar 2BRB, a motor with a bearing outer ring damage BRN, and a motor with an unbalanced mechanical load UNB.
