Aero-engine fault sample generation method based on genetic algorithm

Abstract

The application provides an aero-engine fault sample generation method based on a genetic algorithm, and belongs to the technical field of aero-engines. The method comprises the following steps: obtaining a reference sample of an engine fault, randomly intercepting two parent sample segments and performing binary coding; expanding the coded parent sample based on the genetic algorithm, and decoding and restoring the sample to obtain a generated sample; calculating a time-domain-based cosine similarity error and a frequency-domain-based spectrum error for the generated sample and the reference sample; obtaining a final cost function of the generated sample based on the two errors; evaluating the quality of the current generated sample based on the final cost function, and determining that the current generated sample is a qualified sample when a preset quality requirement is met; and repeating the above steps until the total number of qualified samples reaches a target sample quantity. The generated data of the scheme can effectively extract fault feature information in the original data, effectively improve the adaptability of the generated sample to the intelligent fault diagnosis algorithm training, and has higher calculation efficiency.

Description

Technical Field

[0001] This application relates to the field of aero-engine technology, and in particular to a method for generating aero-engine fault samples based on a genetic algorithm. Background Technology

[0002] As the core power component of aircraft, aero-engines, including their blades and bearings, often operate under harsh conditions of high temperature, high pressure, and high speed, making these components the most prone to failure. Fault early warning and health management of aero-engines have become a very popular research topic in recent years. Compared with traditional model-based and signal analysis-based fault diagnosis methods, data-driven intelligent fault diagnosis algorithms do not rely on precise mathematical models, offering high diagnostic accuracy and strong scalability. However, current intelligent fault diagnosis algorithms face two main problems in practical applications: 1) Intelligent fault diagnosis algorithms often require a large amount of fault data for training, while practical applications typically only provide a small amount of experimental fault data; 2) The imbalance of fault data: Aero-engines experience various faults during operation, such as blade cracks and bearing wear. The experimental data for these various faults is not uniform in size; for example, the number of blade faults may far exceed the number of bearing faults. This imbalanced fault dataset reduces the fault classification accuracy of intelligent fault diagnosis algorithms.

[0003] To address the aforementioned issues, a common approach is to increase the total amount of fault experiment data or the sample size for specific fault types through data augmentation. There are three main existing data augmentation methods:

[0004] 1) Data augmentation based on image processing This method treats the data samples to be augmented as images. It defines operations such as stretching, rotation, translation, and noise reduction, along with parameter sets for each operation. These operations are then randomly combined and applied to the original data to obtain new data samples. Repeating this process multiple times achieves data augmentation. While simple and easy to implement, the generated data exhibits significant randomness, and defining the parameter sets requires some expertise or repeated trial and error to find suitable values.

[0005] 2) Data augmentation based on upsampling or resampling This method expands the data by reducing the sampling interval or performing multiple random samplings of the same signal. While simple, the newly generated data comes entirely from the original data itself, introducing no new information and resulting in information redundancy, which is detrimental to training intelligent fault diagnosis algorithms.

[0006] 3) Data augmentation based on Generative Adversarial Networks (GANs) Generative Adversarial Networks (GANs) are deep learning networks consisting of a generator and a discriminator, commonly used for generating data or images. The generator is responsible for generating a signal with the same dimensions as the original signal from a random signal, while the discriminator compares the similarity between the generated and original signals. GANs are applicable to various data distributions and generate data with high accuracy, leading to their increasingly widespread use in recent years. However, training GANs is very time-consuming and convergence is difficult.

[0007] In summary, while the above methods can generate the massive amounts of data needed for algorithm training or testing based on a small amount of experimental data, and the generated data has a high degree of similarity to the original signal overall, they have the following problems: they do not consider the similarity between the generated signal and the original signal in the frequency domain, while features in the frequency domain often contain more fault information and can more directly reflect the fault state of the system, which is also a more important feature for data-driven fault diagnosis algorithms. Summary of the Invention

[0008] In view of this, embodiments of this application provide a method for generating aero-engine fault samples based on genetic algorithms, which at least partially solves the problem of generating high-quality fault data in the prior art. This method combines genetic algorithms (GA) with fault mechanisms to propose a mechanism-guided GA. Its core idea is to consider the fault mechanism characteristics of aero-engines in the cost function of GA to guide the generation of new samples that have high similarity to reference samples in both the time and frequency domains.

[0009] This application provides a method for generating aero-engine fault samples based on a genetic algorithm, including: Step 1: Obtain a reference sample of engine faults; Step 2: Randomly select two parent samples from the reference sample; Step 3: Encode the two parent samples into binary format; Step 4: Based on the genetic algorithm, the two encoded parent samples are expanded to obtain the encoded expanded samples, and the expanded samples are decoded and restored to obtain the generated samples. Step 5: Calculate the time-domain cosine similarity error and the frequency-domain spectral error for the current generated sample and the reference sample; Step 6: Based on the cosine similarity error and the spectral error, obtain the final cost function corresponding to the current generated sample; Step 7: Evaluate the quality of the current generated sample based on the final cost function. When the preset quality requirements are met, the current generated sample is determined to be a qualified sample. Step 8: Repeat steps 2 to 7 until the total number of qualified samples reaches the target number of samples, and output all qualified samples.

[0010] According to a specific implementation of an embodiment of this application, the binary encoding of the two parent samples includes: For each parent sample, the value of each sampling point in the parent sample is normalized to the range of [-1, 1] to obtain a normalized sample; Based on the required accuracy after sample normalization, determine the number of bits required for encoding each sampling point in the normalized sample; Based on the number of bits in the encoding, each sampling point in the normalized sample is discretized to obtain a discretized sample; Discretized samples are binary encoded according to the number of bits used for encoding; The binary codes of all sampling points are concatenated to obtain the encoding sequence of the parent sample.

[0011] According to a specific implementation of an embodiment of this application, the normalization calculation formula is as follows: , Where x is the value of the current sampling point, min(x) is the minimum value of all sampling points in the parent sample, and max(x) is the maximum value of all sampling points in the parent sample. The value is the normalized value; The formula for calculating the number of bits in the encoding is: , Where f0 is the precision requirement and S is the number of bits used for encoding; The formula for calculating the discretization is: , Where, x D The discretized value, This is the floor operator.

[0012] According to a specific implementation of an embodiment of this application, the step of expanding the encoded two parent samples based on a genetic algorithm includes: Based on the crossover and / or mutation operations in the genetic algorithm, the two encoded parent samples are expanded.

[0013] According to a specific implementation of an embodiment of this application, the crossover operation includes: The two encoded parent samples are treated as two chromosomes; A crossover point is randomly selected on two chromosomes as a hybridization site, and genes in the part before or after the hybridization site are exchanged. Based on the hybridization sites and the identified exchange portions, the genes of the two chromosomes are exchanged and recombinated to obtain an augmented sample with coding.

[0014] According to a specific implementation of an embodiment of this application, the mutation operation includes: The two encoded parent samples are treated as two chromosomes; Set the preset mutation probability; Based on the preset mutation probability, calculate the total number N of gene loci requiring mutation operations. m , , P is the floor operator. m The preset mutation probability is N0, which is the length of the encoded parent sample. Randomly select N from each chromosome. m Each gene locus is mutated to obtain an augmented sample with coding.

[0015] According to a specific implementation of an embodiment of this application, the expression for the cosine similarity error is: , Where cosθ is the cosine similarity error, x gen To generate samples, x ref For reference sample, Norm operators; The expression for the spectral error is: , Among them, L frequency For the spectral error, N f The maximum order of the fault characteristic frequency. and Let be the amplitudes of the generated sample and the reference sample at the i-th order fault characteristic frequency, respectively. and These are the i-th order fault feature frequencies of the generated sample and the reference sample, respectively.

[0016] According to a specific implementation of an embodiment of this application, the expression of the final cost function is: , Among them, L final The final cost function is α, where α is the weighting coefficient for temporal similarity and β is the weighting coefficient for frequency similarity.

[0017] According to a specific implementation of an embodiment of this application, the preset quality requirement is as follows: The final cost function is less than or equal to the preset target value.

[0018] According to a specific implementation of an embodiment of this application, the step of randomly selecting two parent samples from the reference sample includes: Determine the sample length L, and randomly extract two segments of signal with length L from the reference sample as two parent samples.

[0019] Beneficial effects: Existing data augmentation algorithms primarily focus on the overall similarity between the newly generated signal and the original signal in a given single dimension. For example, if the original signal is in the time domain, the newly generated signal is only close to the original signal in the time domain, and the focus is only on the overall similarity of the signals. For fault diagnosis, in addition to the overall similarity of the signal in the time domain, the similarity of the signal in the frequency domain, especially the similarity at specific fault feature frequencies, can more accurately and comprehensively reflect the quality of fault data generation. The aero-engine fault sample generation method based on genetic algorithm proposed in this application uses the sum of the amplitude and frequency errors between the generated data and the original data at specific fault feature frequencies (such as the first 5 fault feature frequencies) as a correction term in the GA cost function. This ensures that the generated data can effectively extract fault feature information from the original data, so that the final generated fault sample can maintain the overall similarity with the original reference sample in the time domain and retain key fault feature information in the frequency domain. This effectively improves the adaptability of the generated sample to the training of intelligent fault diagnosis algorithms, solves the problem of insufficient fault diagnosis accuracy in small sample and imbalanced sample scenarios, and at the same time, this method does not require complex network training like generative adversarial networks, resulting in faster convergence speed and higher computational efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a method for generating aero-engine fault samples based on a genetic algorithm according to an embodiment of the present invention; Figure 2 This is a system overall block diagram according to an embodiment of the present invention; Figure 3 This is an example diagram of sample encoding according to an embodiment of the present invention; Figure 4 This is a schematic diagram of chromosomal gene mutation according to an embodiment of the present invention; Figure 5 This is a time-domain comparison diagram of a generated sample and a reference sample according to an embodiment of the present invention; Figure 6This is a frequency domain comparison diagram of a generated sample and a reference sample according to an embodiment of the present invention. Detailed Implementation

[0022] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0023] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0025] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The illustrations only show the components related to this application and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0026] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0027] This application provides a method for generating aircraft engine fault samples based on a genetic algorithm, which will be described in detail below with reference to the accompanying drawings.

[0028] In one embodiment, a method for generating aero-engine fault samples based on a genetic algorithm is provided, referring to... Figure 1 This includes the following steps: Step 1: Obtain a reference sample of engine faults; Step 2: Randomly select two parent samples from the reference sample; Step 3: Encode the two parent samples into binary format; Step 4: Based on the genetic algorithm, the two encoded parent samples are expanded to obtain the encoded expanded samples, and the expanded samples are decoded and restored to obtain the generated samples. Step 5: Calculate the time-domain cosine similarity error and the frequency-domain spectral error for the current generated sample and the reference sample; Step 6: Based on the cosine similarity error and the spectral error, obtain the final cost function corresponding to the current generated sample; Step 7: Evaluate the quality of the current generated sample based on the final cost function. When the preset quality requirements are met, the current generated sample is determined to be a qualified sample. Step 8: Repeat steps 2 to 7 until the total number of qualified samples reaches the target number of samples, and output all qualified samples.

[0029] The proposed aero-engine fault sample generation method based on genetic algorithm in this embodiment uses the sum of the amplitude and frequency errors between the generated data (generated samples) and the original data (reference samples) at specific fault feature frequencies (such as the first 5 fault feature frequencies) as a correction term in the GA cost function. This ensures that the generated data can effectively extract fault feature information from the original data, so that the final generated fault samples can maintain overall similarity with the original reference samples in the time domain and retain key fault feature information in the frequency domain. This effectively improves the adaptability of the generated samples to the training of intelligent fault diagnosis algorithms, solves the problem of insufficient fault diagnosis accuracy in scenarios with small samples and imbalanced samples, and at the same time, this method does not require complex network training like generative adversarial networks, resulting in faster convergence speed and higher computational efficiency.

[0030] like Figure 2 The system block diagram shown illustrates that, based on the provided reference fault samples, the mechanism-guided GA designed in this application can generate new, high-quality fault samples. Specifically, the mechanism-guided GA makes two modifications to the cost function of commonly used GAs: one is the cosine similarity error between the generated sample and the reference sample in the time domain, and the other is the spectral error between the generated sample and the reference sample in the frequency domain. A detailed description will follow with reference to the accompanying figures.

[0031] In one embodiment, obtaining a reference sample of engine faults includes: Obtain the necessary parameters for the algorithm, including: reference sample x. ref Fault characteristic frequency F of the reference sample ref and the number of target samples N min The reference sample can be a single sample or multiple samples.

[0032] Furthermore, the step of randomly selecting two parent samples from the reference sample includes: Determine the sample length L, and randomly extract two segments of signal with length L from the reference sample as two parent samples.

[0033] For example, in data-driven fault diagnosis, the sample length is usually defined as L=1024, meaning that the sample contains 1024 signal parameters.

[0034] In one embodiment, converting the selected parent sample signal into 0 / 1 binary code includes five steps: normalization, determining the number of bits for encoding, discretization, encoding, and concatenation. Specifically, the binary encoding of the two parent samples includes: Step 3.1, Normalization: For each parent sample, normalize the value of each sampling point in the parent sample to the range of [-1, 1] to obtain normalized samples; Step 3.2: Determine the number of bits required for encoding: Based on the required accuracy after sample normalization, determine the number of bits required for encoding each sampling point in the normalized sample; Step 3.3, Discretization: Based on the number of bits in the encoding, each sampling point in the normalized sample is discretized to obtain a discretized sample; Step 3.4, Encoding: Encode the discretized samples in binary according to the number of bits required; Step 3.5, Concatenation: Concatenate the binary codes of all sampling points to obtain the encoding sequence of the parent sample.

[0035] Furthermore, normalization refers to using a min-max scaling method to normalize the value of the initial random signal to the range [-1, 1]. The normalization calculation formula is as follows: , Where x is the value of the current sampling point, min(x) is the minimum value of all sampling points in the parent sample, and max(x) is the maximum value of all sampling points in the parent sample. The value is the normalized value; The formula for calculating the number of bits in the encoding is: , Where f0 is the precision requirement and S is the number of bits in the encoding; for example, when S is 8, the precision requirement after sample normalization is approximately 0.0078125.

[0036] The formula for calculating the discretization is: , Where, x D The discretized value, This is the floor operator.

[0037] In practice, the signal after direct normalization may contain multiple decimal places instead of integer values, making it inconvenient to encode. Therefore, before encoding, the normalized signal needs to be discretized. Taking 8-bit encoding as an example, its encoding range is [0, 255]. Therefore, the minimum value of the signal, -1, can be mapped to 0, the maximum value of the signal, 1, can be mapped to 255, and other signal values ​​can be mapped to integer values ​​between 0 and 255 using a proportional method.

[0038] Based on step 3.3, each sampling point has been discretized into an integer between 0 and 255. Therefore, it can be encoded using a sequence of 0s and 1s. For example, if the value of a sample point after discretization is 164, its encoding sequence is 10100100.

[0039] like Figure 3 As shown, starting from the first sampling point of the sample, a sequence of length S consisting of 0s and 1s can be generated according to steps 3.1 to 3.4 above. This sequence is the 0 / 1 binary code of that sampling point. Then, similarly, the 0 / 1 encoding is performed on the 2nd to Lth sampling points. Then, the codes of all sampling points are concatenated to obtain the encoding sequence of the entire signal sample, with a sequence length of N0 = L × S. This sequence length is the length of the encoded parent sample.

[0040] In one embodiment, the augmentation of the two encoded parent samples based on a genetic algorithm includes: Based on the crossover and / or mutation operations in the genetic algorithm, the two encoded parent samples are expanded.

[0041] In one embodiment, crossover is the process of recombining and redistributing genes from two paternal chromosomes to produce the next generation. Common hybridization methods include single-point hybridization, two-point hybridization, and uniform hybridization. This embodiment uses single-point hybridization as an example, and the crossover operation includes: The two encoded parent samples are treated as two chromosomes; A crossover point is randomly selected on two chromosomes as a hybridization site, and genes in the part before or after the hybridization site are exchanged. Based on the hybridization sites and the identified exchange portions, the genes of the two chromosomes are exchanged and recombinated to obtain an augmented sample with coding.

[0042] For example, if the two chromosome sequences of the parent are 10010|111 and 00101|010, where "|" represents the crossover point, and the chromosomes of the next generation are 00101|111 and 10010|010 respectively, then the chromosomes of the next generation will be 00101|111 and 10010|010 respectively.

[0043] In one embodiment, besides hybridization, mutation is also a way to generate new genes in a genetic algorithm. Therefore, based on the next generation of chromosomes formed by hybridization, gene mutation operations can be performed with a certain probability, including: The two encoded parent samples are treated as two chromosomes; Set a preset mutation probability according to actual needs. This mutation probability represents the probability of chromosome mutation when generating new offspring individuals. It is usually set to 1% or less. Based on the preset mutation probability, calculate the total number N of gene loci requiring mutation operations. m , , P is the floor operator. m The preset mutation probability is N0, which is the length of the encoded parent sample. Randomly select N from each chromosome. m Each gene locus, as follows: Figure 4 The mutation operation shown refers to mutating the gene at the selected location, changing 1 to 0 or 0 to 1, to obtain an augmented sample with coding.

[0044] In one embodiment, decoding and restoration involves restoring the 0 / 1 encoded samples generated through hybridization and mutation operations into physically meaningful sample signals. Specifically, this includes decoding, dediscretization, and denormalization—that is, the reverse operation of step 3 above. Decoding and restoration specifically includes: Step 4.1: Separate the encoded sequence: The final encoded sequence generated by the above operations has a length of N0 = L × S, which is then separated into L independent codes of length S.

[0045] Step 4.2, Reverse Encoding: The separated individual codes are then restored to their specific numerical values. For example, if the individual code for a sample point is 10100100, its restored value is 164. Using the same method, the inverse encoding can be completed for L sample points contained in the sample.

[0046] Step 4.3, Dediscretization: Each sampled signal after de-encoding is de-discretized. Taking 8-bit encoding as an example, its encoding range is [0, 255]. Therefore, the minimum value of the signal 0 can be mapped to -1, the maximum value of the signal 255 can be mapped to 1, and the values ​​of other signals are mapped to values ​​between [-1, 1] according to a proportional method. For example: the de-encoded value x D If '=164, then the value after dediscretization =0.286.

[0047] The calculation method is as follows: .

[0048] Step 4.4, Inverse Normalization: The signal at each sample point after dediscretization is denormalized to form a sample with physical meaning. The specific calculation formula is as follows: .

[0049] In this embodiment, the purpose of decoding and restoring the expanded samples is to convert the binary code back into actual engine vibration signal parameters, which facilitates subsequent error calculation and sample quality evaluation, and ensures that the generated samples can be directly used for training the fault diagnosis model.

[0050] In one embodiment, in order to evaluate the generated sample (x) gen ) and reference sample (x) ref The similarity between the two is evaluated using cosine similarity as a quantitative indicator in this embodiment. The expression for the cosine similarity error is as follows: , Where cosθ is the cosine similarity error, x gen To generate samples, x ref For reference sample, Norm operators; In addition to needing high similarity to the reference sample in the time domain, the generated sample also needs to have as high a similarity as possible to the reference sample in the spectrum, especially at the fault characteristic frequency F. ref The amplitude and frequency errors under its higher harmonics are investigated, and the following expression for the spectral error is proposed: , Among them, L frequency For the spectral error, N f The maximum order of the fault characteristic frequency can be N. f =3, and Let be the amplitudes of the generated sample and the reference sample at the i-th order fault characteristic frequency, respectively. and These are the i-th order fault feature frequencies of the generated sample and the reference sample, respectively.

[0051] In one embodiment, to simultaneously ensure the similarity between the generated sample and the reference sample in both the time and frequency domains, this embodiment uses the weighted sum of the cosine similarity error and the spectral error as the final cost function. The expression for the final cost function is as follows: , Among them, L final The final cost function is defined by α, which is the weighting coefficient for temporal similarity, and β, which is the weighting coefficient for frequency similarity. α and β can be adjusted by the user. It is generally recommended to satisfy α+β=1 for normalization, for example, α=0.5, β=0.5.

[0052] In one embodiment, the preset quality requirement is: The final cost function is less than or equal to the preset target value.

[0053] In practice, the quality of each newly generated sample is evaluated based on the final cost function constructed in step 6 above. When its cost function L... final Less than or equal to the preset target value L aim (i.e. L) final ≤L aim When the time reaches t, the generated sample is considered qualified. Repeat steps 2 to 7, and count the total number of qualified samples N generated up to the current time t. t When the total number of qualified samples reaches the target sample size (N) t ≥N min When the sample generation is complete, the process can be terminated.

[0054] Once the samples are generated, the final step is to output all samples that meet the requirements. These newly generated qualified samples can be used to train fault diagnosis algorithms to improve the accuracy of fault diagnosis in practical applications with small or imbalanced datasets.

[0055] Figure 5 and Figure 6 The implementation effects of this application are given, among which, Figure 5 The paper presents a time-domain comparison between the reference sample and the generated sample, and it can be found that the generated sample has a high degree of similarity to the reference sample in the time domain. Figure 6The frequency domain analysis of the reference sample and the generated sample is given separately. The vertical dashed lines in the figure represent the theoretical fault characteristic frequencies of orders 1 to 3. In the spectrum, it can be found that the sample generated in this application has a peak value similar to that of the reference sample, further illustrating that the signal generated by the improved genetic algorithm GA in this application can more effectively extract fault features from the original signal.

[0056] To address the problem of generating high-quality fault data, this application proposes a mechanism-guided GA that combines genetic algorithms (GA) with fault mechanisms. The core idea is to incorporate the fault mechanism characteristics of aero-engines into the cost function of the GA, guiding the generation of new samples that exhibit high similarity to reference samples in both the time and frequency domains. In summary, the application mainly involves the following three aspects: 1) Construct the cost function of GA from the time domain and frequency domain respectively.

[0057] First, to evaluate the temporal similarity between the generated sample and the reference sample, a cosine similarity error is defined: , The cosine similarity error cosθ can reflect the degree of similarity between two signals as a whole.

[0058] Secondly, considering that in practical applications, although the generated signal has a high similarity to the original signal in the time domain, after spectral transformation, the generated signal does not have the same or similar fault characteristic frequency as the original signal, this application uses the sum of the amplitude and frequency errors of the envelope spectrum of the generated signal and the original signal at the fault characteristic frequency as another correction term of the GA cost function, as shown in the following equation: , Where, N f This is the maximum order of the fault characteristic frequency, which can be set according to engineering application requirements. and Let be the amplitudes of the generated sample and the reference sample at the i-th order fault characteristic frequency, respectively. and These are the i-th order fault feature frequencies of the generated sample and the reference sample, respectively.

[0059] The selection of fault characteristic frequencies depends on the fault category of the original signal. When different components of an aero-engine malfunction, corresponding fault characteristic frequencies and their harmonics will appear in the envelope spectrum of the vibration signal. For example, if a blade malfunctions, the corresponding fault characteristic frequency is the Blade Passing Frequency (BPF); if the bearing outer race malfunctions, the corresponding fault characteristic frequency is the Ball Pass Frequency Outer Race (BPFO).

[0060] 2) Integrate temporal error and spectral error to generate the final cost function of GA.

[0061] After considering the two correction terms mentioned above, the final form of the GA cost function is defined as: , Wherein, α is the weighting coefficient of temporal similarity, and β is the weighting coefficient of frequency similarity, which can be dynamically adjusted by the user according to different needs.

[0062] 3) Specify the precision of the sample 0 / 1 binary encoding.

[0063] The original measured signals are generally floating-point variables with both positive and negative signs, making them unsuitable for direct encoding in genetic algorithms. This application proposes a sample encoding method consisting of five steps: normalization, determining the number of bits for encoding, discretization, encoding, and concatenation. This method can achieve 0 / 1 binary encoding of signals at any specified precision.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating aero-engine fault samples based on a genetic algorithm, characterized in that, include: Step 1: Obtain a reference sample of engine faults; Step 2: Randomly select two parent samples from the reference sample; Step 3: Encode the two parent samples into binary format; Step 4: Based on the genetic algorithm, the two encoded parent samples are expanded to obtain the encoded expanded samples, and the expanded samples are decoded and restored to obtain the generated samples. Step 5: Calculate the time-domain cosine similarity error and the frequency-domain spectral error for the current generated sample and the reference sample; Step 6: Based on the cosine similarity error and the spectral error, obtain the final cost function corresponding to the current generated sample; Step 7: Evaluate the quality of the current generated sample based on the final cost function. When the preset quality requirements are met, the current generated sample is determined to be a qualified sample. Step 8: Repeat steps 2 to 7 until the total number of qualified samples reaches the target number of samples, and output all qualified samples.

2. The method for generating aero-engine fault samples based on genetic algorithms according to claim 1, characterized in that, The binary encoding of the two parent samples includes: For each parent sample, the value of each sampling point in the parent sample is normalized to the range of [-1, 1] to obtain a normalized sample; Based on the required accuracy after sample normalization, determine the number of bits required for encoding each sampling point in the normalized sample; Based on the number of bits in the encoding, each sampling point in the normalized sample is discretized to obtain a discretized sample; Discretized samples are binary encoded according to the number of bits used for encoding; The binary codes of all sampling points are concatenated to obtain the encoding sequence of the parent sample.

3. The method for generating aero-engine fault samples based on genetic algorithms according to claim 2, characterized in that, The normalization calculation formula is as follows: ， Where x is the value of the current sampling point, min(x) is the minimum value of all sampling points in the parent sample, and max(x) is the maximum value of all sampling points in the parent sample. The value is the normalized value; The formula for calculating the number of bits in the encoding is: ， Where f0 is the precision requirement and S is the number of bits used for encoding; The formula for calculating the discretization is: ， Where, x D The discretized value, This is the floor operator.

4. The method for generating aero-engine fault samples based on genetic algorithms according to claim 1, characterized in that, The expansion of the two encoded parent samples based on the genetic algorithm includes: Based on the crossover and / or mutation operations in the genetic algorithm, the two encoded parent samples are expanded.

5. The method for generating aero-engine fault samples based on genetic algorithms according to claim 4, characterized in that, The crossover operation includes: The two encoded parent samples are treated as two chromosomes; A crossover point is randomly selected on two chromosomes as a hybridization site, and genes in the part before or after the hybridization site are exchanged. Based on the hybridization sites and the identified exchange portions, the genes of the two chromosomes are exchanged and recombinated to obtain an augmented sample with coding.

6. The method for generating aero-engine fault samples based on a genetic algorithm according to claim 4, characterized in that, The mutation operation includes: The two encoded parent samples are treated as two chromosomes; Set the preset mutation probability; Based on the preset mutation probability, calculate the total number N of gene loci requiring mutation operations. m , , P is the floor operator. m The preset mutation probability is N0, which is the length of the encoded parent sample. Randomly select N from each chromosome. m Each gene locus is mutated to obtain an augmented sample with coding.

7. The method for generating aero-engine fault samples based on genetic algorithm according to claim 1, characterized in that, The expression for the cosine similarity error is: ， Where cosθ is the cosine similarity error, x gen To generate samples, x ref For reference sample, Norm operators; The expression for the spectral error is: ， Among them, L frequency For the spectral error, N f The maximum order of the fault characteristic frequency. and Let be the amplitudes of the generated sample and the reference sample at the i-th order fault characteristic frequency, respectively. and These are the i-th order fault feature frequencies of the generated sample and the reference sample, respectively.

8. The method for generating aero-engine fault samples based on a genetic algorithm according to claim 7, characterized in that, The final cost function is expressed as follows: ， Among them, L final The final cost function is α, where α is the weighting coefficient for temporal similarity and β is the weighting coefficient for frequency similarity.

9. The method for generating aero-engine fault samples based on genetic algorithms according to claim 8, characterized in that, The preset quality requirements are as follows: The final cost function is less than or equal to the preset target value.

10. The method for generating aero-engine fault samples based on a genetic algorithm according to claim 1, characterized in that, The step of randomly selecting two parent samples from the reference sample includes: Determine the sample length L, and randomly extract two segments of signal with length L from the reference sample as two parent samples.

Images