Adaptive sampling method for fault diagnosis under multi-class imbalance of data and related equipment

Abstract

The invention provides a self-adaptive sampling method for fault diagnosis under data multi-class imbalance and related equipment, and effectively solves the problems of low sample generation quality, high parameter dependence, high calculation complexity, poor adaptability to complex working conditions and the like. The method comprises the steps of coping with different data feature scenes through a parameter adaptive calculation mechanism, then exploring a global optimal solution in multi-classification modeling accuracy model solution optimization by an evolution mechanism through employing a Newton-Raphson optimizer thought, and finally converting an optimal solution set into various fault samples by using a feature recombination mechanism. Therefore, high-quality sample equalization is realized, small sample multi-class imbalance fault diagnosis is carried out by combining MAESTE and multi-class LS-SVM, and the interpretability of the model is improved.

Description

Technical Field

[0001] This application relates to the technical field of mechanical fault diagnosis, and particularly to an adaptive sampling method for fault diagnosis under multi-class data imbalance and related devices. Background Art

[0002] With the rapid development of the big data era, the importance of mechanical equipment health monitoring and fault diagnosis technology in high-end manufacturing has become increasingly prominent. By analyzing a large amount of equipment operation data, potential faults can be identified in a timely manner, accidental equipment shutdowns can be prevented, production efficiency can be improved, and operation safety can be ensured. However, in actual industrial applications, due to the diverse and complex fault types, often accompanied by simultaneous faults of multiple components or multiple fault modes of the same component, and at the same time, the normal state data of the equipment is much more than the fault state data, and the fault samples often have characteristics such as uneven distribution and scarcity in quantity, forming a typical small sample multi-class imbalance data feature, which brings great challenges to the intelligent fault diagnosis of high-end equipment.

[0003] Currently, four main methods are commonly used for fault diagnosis under multi-class imbalance data: data sampling technology, classifier adaptation technology, ensemble learning method, and cost-sensitive method. Among them, traditional sampling technology balances the class distribution by interpolating and synthesizing or replicating minority class samples. This type of method is easy to implement, but it is difficult to control the quality of the synthesized samples, and it may introduce noise or lead to overfitting, especially showing instability in multi-class complex scenarios; classifier adaptation technology enhances the model's recognition ability for minority classes by modifying the classifier structure or loss function (such as Focal Loss, multi-task learning, etc.). Although the model performance is improved to a certain extent, this type of method is sensitive to hyperparameters, complex to debug, and has limited effects when the data classes are highly imbalanced. Ensemble learning methods improve the model's generalization ability by combining multiple base learners. This method can alleviate the data imbalance problem to a certain extent, but has a high computational cost, a long model training time, and limited performance improvement when the samples are extremely scarce. The cost-sensitive method enhances the model's attention to minority class samples by introducing class weights or misclassification costs during the training process. This type of method does not rely on sample generation, but requires accurate setting of cost weights and has poor adaptability to different data sets, facing great uncertainties in practical applications.

[0004] In summary, in the field of high-end equipment fault diagnosis research under actual working conditions, whether it is the data-level processing method or the technology-level technology, although it alleviates the problem of the decline in diagnostic performance caused by small sample multi-class imbalance data to a certain extent, there are generally problems such as low sample generation quality, strong parameter dependence, high computational complexity, and poor adaptability to complex working conditions. Therefore, there is an urgent need for a more efficient, stable, and adaptive diagnostic method to solve the above problems. Summary of the Invention

[0005] The present application provides an adaptive sampling method and related devices for fault diagnosis under multi-class data imbalance to solve the above technical problems.

[0006] The first aspect of the present application provides an adaptive sampling method for fault diagnosis under multi-class data imbalance. The method at least includes the following steps: Divide each category of samples according to the original dataset, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method for denoising preprocessing; Perform feature processing on the denoised dataset. Use the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data to obtain a series of intrinsic mode functions, calculate the Shannon entropy for each modal function, and extract its eigenvalue to obtain the reconstructed feature dataset; According to the analysis of the data distribution characteristics, calculate the sample imbalance ratio and the sample difference information within and between classes of different datasets, and dynamically adjust the key parameters in the sampling process; By defining an objective function that minimizes the classification error, adopt the Newton-Raphson evolutionary mechanism to obtain the global optimal solution in the solution optimization of the multi-classification modeling accuracy model. The Newton-Raphson evolutionary mechanism includes the Newton-Raphson search and the trap avoidance mechanism; Use the feature recombination mechanism to transform the generated optimal solution samples into new synthesized samples of the minority class, and adopt the fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnostic model; Adopt the MAESTE method combined with multi-class LS-SVM for fault diagnosis of small-sample multi-class imbalanced data.

[0007] Optionally, dividing each category of samples according to the original dataset, calculating the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and using the k-nearest neighbor method for denoising preprocessing includes: Divide the original dataset into majority-class samples and minority-class samples according to the number of samples in each category, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method to perform denoising preprocessing on each category of samples.

[0008] Optionally, calculating the Shannon entropy for each modal function and extracting its eigenvalue to obtain the reconstructed feature dataset includes: Adopt the ICEEMDAN-Shannon energy entropy method to perform feature extraction and reconstruct the dataset on the data after denoising preprocessing.

[0009] Optionally, according to the analysis of the data distribution characteristics, calculating the sample imbalance ratio and the sample difference information within and between classes of different datasets, and dynamically adjusting the key parameters in the sampling process includes: Dynamically and adaptively calculate the number of fault samples to be generated according to different data characteristics, and then calculate the total number of features to be solved in the solution process according to the total number of samples to be generated and the number of features.

[0010] Optionally, by defining an objective function that minimizes the classification error, and using the Newton-Raphson evolutionary mechanism to obtain the global optimal solution in the solution optimization of the multi-class modeling accuracy model, the Newton-Raphson evolutionary mechanism includes Newton-Raphson search and trap avoidance mechanisms, including: Define the objective function in the solution process to obtain the minimum classification error, and use the Newton-Raphson search criterion to explore and solve the global optimal solution.

[0011] Optionally, use the feature recombination mechanism to transform the generated optimal solution samples into new synthesized samples of the minority class, and use the fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnostic model, including: Use the feature recombination mechanism to recombine the optimal solution to generate minority class fault samples; Perform fitness evaluation on the generated samples to ensure that the quality of the newly generated samples meets the requirements; Use the fitness evaluation mechanism to evaluate by measuring the performance of the new samples in the classification model. If the fitness of the newly generated samples is better, update the samples. The expression of the fitness function is as follows:

[0012] Where ACC is the classification error, G represents the error of the G-Mean index, F represents the F-Measure error, A represents the AUC error, and 𝛼1 and 𝛼2 are two weight parameters.

[0013] Optionally, combine MAESTE with multi-class LS-SVM for fault diagnosis of small sample multi-class imbalanced data, including: Merge the newly generated high-quality samples with the original sample set, and divide them into a training sample set and a test sample set. Train multiple LS-SVM models using the one-versus-one strategy and classify through a voting mechanism; Use the test sample set as the input of the multi-class LS-SVM, calculate four evaluation indicators of accuracy, G-mean, F-measure, and AUC, and evaluate the MAESTE method through the four evaluation indicators; Combine the MAESTE method with multi-class LS-SVM for fault diagnosis of small sample multi-class imbalanced data.

[0014] The second aspect of this application provides an adaptive sampling system for fault diagnosis under multi-class imbalance of data. The system includes: Partitioning unit, configured to partition each category of samples according to the original data set, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method for denoising preprocessing; Processing unit, configured to perform feature processing on the denoised data set, use the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data, obtain a series of intrinsic mode functions, calculate the Shannon entropy for each modal function, and extract its eigenvalue to obtain the reconstructed feature data set; Adjusting unit, configured to calculate the sample imbalance ratio and the sample difference information between intra-class and inter-class of different data sets according to the analysis of data distribution characteristics, and dynamically adjust the key parameters in the sampling process; Obtaining unit, configured to obtain the global optimal solution in the multi-classification modeling accuracy model solution optimization by defining an objective function that minimizes the classification error and adopting the Newton-Raphson evolutionary mechanism, where the Newton-Raphson evolutionary mechanism includes the Newton-Raphson search and the trap avoidance mechanism; Evaluating unit, configured to use the feature recombination mechanism to transform the generated optimal solution samples into new synthetic samples of the minority class, and use the fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnostic model; Diagnostic unit, configured to perform fault diagnosis on small-sample multi-class imbalanced data by combining the MAESTE method with multi-class LS-SVM The third aspect of the present application provides an adaptive sampling device for fault diagnosis under multi-class data imbalance, and the device includes: A processor, a storage, an input / output unit, and a bus; The processor is connected to the storage, the input / output unit, and the bus; The storage stores a program, and the processor calls the program to execute the method of the first aspect and any optional method in the first aspect.

[0015] The fourth aspect of the present application provides a computer-readable storage medium, and a program is stored on the computer-readable storage medium, and when the program is executed on a computer, it executes the method of the first aspect and any optional method in the first aspect.

[0016] It can be seen from the above technical solutions that the present application has the following advantages: This application uses a parameter adaptive calculation mechanism to cope with different data feature scenarios. Then, the evolutionary mechanism explores the global optimal solution in the solution optimization of the multi-classification modeling accuracy model by employing the idea of the Newton-Raphson optimizer. Next, the feature recombination mechanism is used to transform the optimal solution set into various fault samples, thereby achieving high-quality sample balance. Finally, combined with multi-class LS-SVM for classification modeling to improve the model interpretability, in order to effectively improve the recognition ability of the classification model for various fault samples in the complex data.

[0017] Therefore, this application effectively solves problems such as low quality of sample generation, strong parameter dependence, high computational complexity, and poor adaptability to complex working conditions. MAESTE cleverly solves the problem of fault diagnosis modeling for high-end equipment such as tool and bearing systems under small sample multi-class unbalanced data, that is, the limitations of the existing sampling technology, such as strong parameter dependence, uncontrollable quality of generated samples, and inability to meet unknown complex scenarios, and has a wider applicable scenario. Brief Description of the Drawings

[0018] In order to more clearly illustrate the technical solutions in this application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0019] Figure 1 It is a schematic flowchart of an embodiment of the adaptive sampling method for fault diagnosis under multi-class unbalanced data of this application; Figure 2 It is a schematic diagram of an embodiment of the adaptive sampling system for fault diagnosis under multi-class unbalanced data of this application; Figure 3 It is a schematic diagram of an embodiment of the adaptive sampling device for fault diagnosis under multi-class unbalanced data of this application; Figure 4 It is a schematic diagram of an embodiment of the cooperation relationship between the MAESTE technology and the processing platform of this application; Figure 5 It is a structural diagram of the MAESTE sampling algorithm in this application; Figure 6 It is a framework diagram of the combination of MAESTE and multi-class LS-SVM for small sample multi-class unbalanced fault diagnosis in this application. Detailed Embodiments

[0020] This application provides an adaptive sampling method for fault diagnosis under multi-class unbalanced data and related devices, effectively solving problems such as low quality of sample generation, strong parameter dependence, high computational complexity, and poor adaptability to complex working conditions.

[0021] Please refer to Figure 1 , an embodiment of an adaptive sampling method for fault diagnosis under multi-class data imbalance is provided in the first aspect of this application. This embodiment includes: 101. Divide each category of samples according to the original data set, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method for denoising preprocessing; 102. Perform feature processing on the denoised data set. Use the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data to obtain a series of intrinsic mode functions. Calculate the Shannon entropy for each modal function, extract its eigenvalue to obtain the reconstructed feature data set; 103. According to the analysis of data distribution characteristics, calculate the sample imbalance ratio and the sample difference information between intra-class and inter-class of different data sets, and dynamically adjust the key parameters in the sampling process; 104. By defining an objective function that minimizes the classification error, adopt the Newton-Raphson evolutionary mechanism to obtain the global optimal solution in the solution optimization of the multi-classification modeling accuracy model. The Newton-Raphson evolutionary mechanism includes the Newton-Raphson search and the trap avoidance mechanism; 105. Use the feature recombination mechanism to transform the generated optimal solution samples into new synthesized samples of the minority class, and adopt the fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnostic model; 106. Adopt the MAESTE method combined with multi-class LS-SVM for fault diagnosis of small-sample multi-class imbalanced data.

[0022] In the embodiment of this application, the Newton-Raphson optimizer (NRBO) is a new meta-heuristic optimization algorithm that combines the classical Newton-Raphson method with the swarm intelligence evolutionary strategy. This algorithm utilizes the local search ability of the gradient and Hessian matrix in the Newton-Raphson method (NRSR) to achieve a rapid approximation of the objective function. At the same time, it combines the population search and the trap avoidance mechanism (TAO) to enhance the global search ability and the ability to jump out of the local optimum. Through this optimization method that combines exploration and exploitation, this algorithm shows good convergence speed and solution accuracy in continuous optimization problems, and is particularly suitable for optimization tasks in high-dimensional, non-linear, and complex search spaces.

[0023] The k-NN preprocessing denoising (k-Nearest Neighbors) technique is a method used to clean and optimize data quality before data modeling. It is based on the basic idea of the k-NN algorithm: by analyzing the category or value of each data point and its k nearest neighbors, it determines whether there are significant differences between the data point and its neighbors. If the label of a point is inconsistent with that of most of its neighbors, it may be considered noise and then corrected, smoothed, or removed. This method can effectively reduce the interference of outliers and incorrect labels on model training, thereby improving the classification or regression performance of k-NN and other subsequent algorithms, and is particularly suitable for data sets highly affected by noise.

[0024] The LS-SVM algorithm, namely the Least Squares Support Vector Machines, is a machine learning algorithm used to solve problems such as pattern classification and function estimation. The core of LS-SVM regression lies in selecting an appropriate kernel function. Common kernel functions include linear kernel functions, polynomial kernel functions, Gaussian kernel functions, etc. The choice of kernel function depends on the characteristics of the specific problem. By selecting an appropriate kernel function, LS-SVM can handle nonlinear problems in high-dimensional spaces.

[0025] ICEEMDAN-Shannon feature extraction is a feature extraction method that combines signal decomposition and information entropy analysis. ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) is a signal decomposition technique that can decompose non-stationary signals into a set of intrinsic mode functions (IMFs) with physical meanings, effectively avoiding the problem of mode mixing. Subsequently, by calculating the Shannon entropy of each IMF component, the information complexity and uncertainty in its time series are extracted to form a feature vector reflecting the dynamic characteristics of the signal. This method is widely used in fields such as fault diagnosis and medical signal processing, and has strong time-frequency localization ability and information extraction ability.

[0026] Furthermore, in step 101, when dividing each category of samples according to the original data, in actual working conditions, most data belong to normal operation samples. They are divided into majority class samples, and samples with different faults are respectively divided into minority class samples.

[0027] Even further, in step 101, it includes k-NN data preprocessing, that is, using the k-NN method to identify the noise samples in each category and delete the noise. For a given sample x i , its neighboring sample set is N k ( x i ), then the local density of the sample can be expressed as:

[0028] Among them, dist ( x i , x j ) represents the Euclidean distance between the sample x i and the neighboring samples. For points with significantly lower density than their neighboring samples (i.e., noise points), they are removed from the dataset. x j Furthermore, in step 102, it includes feature processing on the denoised dataset. The ICEEMDAN modal decomposition method is used to perform multi-level decomposition on the vibration signal data to obtain a series of intrinsic mode functions (IMFs). Shannon entropy is calculated for each mode function (IMF), and its eigenvalues are extracted to obtain the reconstructed feature dataset;

[0029] Furthermore, in step 103, it includes calculating the number of generated fault samples. Considering the small sample imbalance problem and the continuity of the sampling process, the following formula is used for dynamic adjustment: Furthermore, in step 103, it includes calculating the number of generated fault samples. Considering the small sample imbalance problem and the continuity of the sampling process, the following formula is used for dynamic adjustment:

[0030] Among them, Num i represents the difference in the number of samples between the i th class of minority class and majority class, Num represents the total number of samples to be generated, N len and P ilen represent the number of normal samples and the number of fault samples in each category. These differences determine the number of samples to be generated. The problem dimension dim is the total number of features to be solved in the solution process, which is determined by the total number of samples to be generated and the number of features. The calculation formula is as follows:

[0031] Among them, fen_num represents the number of features after feature extraction. Therefore, each candidate solution represents a potential sampling strategy, and its length is equal to the product of the total number of samples and the number of features.

[0032] Furthermore, in step 104, it includes defining the objective function. The goal is to find the minimum classification error, and its calculation formula is as follows:

[0033] Among them, S represents the generated samples for training, R prdRepresents the classification accuracy on the test set, is the class label predicted by the model, is the true class label. n represents the total number of test samples, and 1(⋅) is the indicator function, when = its value is 1, otherwise it is 0. Then, during the population initialization process, initial values will be generated for the candidate solutions of each dimension according to the minority class samples, and these values are uniformly randomly distributed between the predetermined upper and lower bounds to ensure that all candidate solutions are within the defined search space. The population initialization formula is as follows:

[0034] where represents the coordinate of the nth individual in the jth dimension, and rand is a random number in the range of 0 to 1. The entire population matrix is obtained as follows:

[0035] Then, the current position vector is updated in position according to the NRSR search criterion, as follows:

[0036] where, ,

[0037] , , r1, r2 represent random numbers between 0 and 1 but they are not equal, and randn represents a random number from a normal distribution with a mean of 0 and a variance of 1, X w is the position vector of the individual with the worst fitness value in the current iteration, X b is the position vector of the current optimal solution, and are the corresponding position updates during the NRSR search process respectively. The parameter guides the population to move in the correct direction and is an adaptive coefficient that controls the balance between exploration and exploitation, ensuring that the algorithm achieves a balance between diversity and intensity.

[0038] Next, the TAO mechanism is used to avoid traps and prevent falling into local optima. The formula is as follows:

[0039] where, , , θ1, and θ2 are random values between 0 and 1, -1 and 1, and -0.5 and 0.5 respectively. β represents a binary number, either 1 or 0, enabling the algorithm to dynamically adjust the search strategy according to the current search state, thereby effectively guiding the population to move towards promising regions while maintaining sufficient exploration ability. When the maximum number of iterations is reached, the algorithm terminates and outputs the optimal solution set.

[0040] Furthermore, in step 105, it includes driving sample generation using a feature recombination mechanism. Each optimal solution corresponds to a part of the feature dimensions, and new samples can be generated by combining these parts. Specifically, each newly generated sample is composed of fen_num optimal solutions with the number of features. This method ensures that while increasing the number of minority class samples, the diversity characteristics of the minority class can be effectively captured. The formula for synthesizing samples can be expressed as:

[0041] where, is the newly generated minority class sample of the i-th class, represents the feature dimension fen _ num corresponding optimal solution. Then, the fitness of the newly generated samples is evaluated to ensure that the quality of these samples meets the requirements of the model. The fitness evaluation is carried out by measuring the performance of the new samples in the classification model. The expression of the fitness function fit is:

[0042] where ACC is the classification error, G represents the error of the G-Mean index, F represents the F-Measure error, A represents the AUC error, 𝛼1 and 𝛼2 are two important weight parameters. The value range of 𝛼1 is between 0 and 1, representing the degree of emphasis on classification accuracy, while 𝛼2 = (1 - 𝛼1) / 3, representing the weights of the other 3 classification evaluation indicators. Considering that the classification accuracy is the main goal of the classifier, therefore, 𝛼1 is set to 0.7 and 𝛼2 is 0.1. If the fitness of the new sample is better, then update the sample position: , by minimizing the fitness function fit , the quality of the generated samples can be ensured, and the recognition ability of the model for minority class samples can be improved.

[0043] Further, in step 106, it includes combining the generated samples and the sample set, and dividing them into a training sample set and a test sample set according to a ratio of 7:3. For the multi-class data classification problem, the one-versus-one strategy is adopted to train multiple LS-SVM models, and the final classification is carried out through a voting mechanism. The test sample set is used as the input of the multi-class LS-SVM, and the label values of each test sample belonging to each class are output. According to the size of the output label values of each class, the category to which the corresponding test sample belongs is judged. Finally, four evaluation indicators, namely accuracy, G-mean, F-measure, and AUC, are calculated to evaluate the MAESTE method. Among them, for the structure of the MAESTE sampling algorithm, the framework diagrams of the combination of MAESTE and multi-class LS-SVM for small sample multi-class unbalanced fault diagnosis can be referred to respectively Figure 5 and Figure 6 。

[0044] To further illustrate the technical solution of the present application, experimental tests were carried out on 3 publicly available data sets and 2 self-collected data according to the actual working conditions. In the following description process, one group of bearing data sets is taken as an example, and the operations of other data sets are the same as those of this data set. The specific implementation plan includes the following contents and steps: Step 1: Divide each category of samples according to the original data. Under actual working conditions, most of the data belongs to normal operation samples, which are divided into majority class samples, and the samples with different faults are respectively divided into minority class samples. Then k-NN is used for data preprocessing and denoising. The specific steps are as follows: Step 1.1: Divide each category of samples according to the original data. The normal operation samples are divided into majority class samples N1, and the samples with different faults are respectively divided into minority class samples P2, P3, P4, and corresponding labels are attached respectively.

[0045] Step 1.2: Use the k-NN method to identify the noise samples in each category and delete the noise. The local density of each category of samples is calculated through the following formula:

[0046] where dist(xi,xj) represents the Euclidean distance between sample xi and neighboring sample xj. For points whose density is significantly lower than that of their neighboring samples (i.e., noise points), they are removed from the data set to obtain the denoised data set.

[0047] Step 2: Perform feature processing on the denoised data set. Use the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data to obtain a series of intrinsic mode functions (IMFs). Calculate the Shannon entropy of each mode function (IMF) and extract its eigenvalue to obtain the reconstructed feature data set.

[0048] Step 3: Adaptively calculate the key parameters of the sampling process according to the data characteristics. The specific steps are as follows: Step 3.1: Dynamically adjust the key parameters according to information such as the sample imbalance ratio and the difference between intra-class and inter-class samples of different datasets. The difference in the number of samples between the majority class and the minority class directly affects the parameter setting. First, it is necessary to calculate the number of generated faulty samples, that is, the difference in the number of samples between the minority class and the majority class. Since the sampling process is continuous, the following formula is used for dynamic adjustment:

[0049] Among them, Numi represents the difference in the number of samples between the minority class and the majority class of the i-th class, that is, the number of samples to be generated, Nlen represents the number of normal multi-class samples, and Pilen represents the number of faulty samples of the i-th class.

[0050] Step 3.1: Next, it is necessary to define two core parameters of the sampling algorithm: population size (pop_size) and problem dimension (dim). The population size pop_size represents the number of candidate solutions generated in the population, and the problem dimension is determined by the total number of samples to be generated and the number of features. The specific formula is:

[0051] Among them, fen_num represents the number of features. Therefore, each candidate solution represents a potential sampling strategy, and its length is equal to the product of the total number of samples and the number of features.

[0052] Step 4: Use the evolutionary mechanism to explore the global optimal solution in the multi-classification modeling accuracy model solving and optimization. The NRBO algorithm combines the advantages of gradient and population optimization methods. By introducing the Newton-Raphson search rule (NRSR) and the trap avoidance operator (TAO), it achieves a balance between exploration and exploitation, thus effectively improving the convergence speed and avoiding local optimal traps. The specific steps are as follows: Step 4.1: Define the objective function to be solved. The goal is to find the minimum classification error:

[0053] Step 4.2: Population initialization process. Generate initial values for the candidate solutions of each dimension according to the minority class samples. These values are uniformly randomly distributed between the predetermined upper and lower boundaries to ensure that all candidate solutions are within the defined search space. The initialization formula is as follows:

[0054] The entire population matrix is obtained as follows:

[0055] Step 4.3: Use the NRSR search criterion to guide the search direction and perform corresponding position updates in NRSR. The expression is as follows: Step 4.4: Using TAO can significantly change the position, avoid falling into local optima. By combining the best position Xb and the current vector position

[0056] a solution with improved quality is obtained. The formula for TAO is: to avoid being trapped in local optima. By combining the best position Xb and the current vector position a solution with improved quality is obtained. The formula for TAO is: to avoid being trapped in local optima. By combining the best position Xb and the current vector position

[0057] When the maximum number of iterations is reached, terminate the algorithm and output the global optimal solution set.

[0058] Step 5: Convert the optimal solution of the objective function into a few - class fault samples through a feature recombination mechanism and evaluate the fitness function of the samples. The specific steps are as follows: Step 5.1: Convert the optimal candidate solutions into new few - class samples through a feature recombination mechanism. Each optimal solution corresponds to a part of the feature dimensions. By combining these parts, new samples can be generated. Specifically, fen_num is the feature dimension, and each generated new sample is composed of fen_num optimal solutions. This method ensures that the generated samples can effectively capture the diversity characteristics of the few - class while increasing the number of few - class samples. The formula for synthesizing samples can be expressed as;

[0059] Step 5.2: After generating new samples, it is necessary to evaluate their fitness to ensure that the quality of these samples meets the requirements of the model. The fitness evaluation is carried out by measuring the performance of the new samples in the classification model. The expression of the fitness function fit is:

[0060] Step 6: Combine with multi - class LS - SVM for small - sample multi - class imbalanced fault diagnosis to increase the interpretability of the model. For the multi - class data classification problem, adopt the one - against - one strategy to train multiple LS - SVM models and perform final classification through a voting mechanism. The specific steps are as follows: Step 6.1: Combine the generated samples and the sample set and divide them into a training sample set and a test sample set, where the training sample set accounts for 70% and the test sample set accounts for 30%; Step 6.2: Train an LS - SVM classifier for each pair of classes, and a total of C(C - 1) / 2 classifiers are trained, where C is the total number of classes; Step 6.3: Use the test sample set as the input of LS-SVM. For each test sample, all classifiers vote, and select the class with the most votes as the final classification result; Step 6.4: Compare the predicted values output by the classification model voting with the actual classes to obtain a confusion matrix, and then calculate the relevant evaluation indicators of the classification model to comprehensively evaluate the samples generated by evolutionary sampling and the trained classification model; Step 6.5: Calculate four evaluation indicators: accuracy, G-mean, F-measure, and AUC to evaluate the MAESTE method.

[0061] To verify the effectiveness and superiority of the proposed MAESTE algorithm, it is evaluated on 3 publicly available datasets and 2 self-collected datasets, and compared with 8 other popular oversampling methods: 1) No-sampling, 2) SMOTE, 3) BSMOTE, 4) IA-SUWO, 5) Cluster-SMOTE, 6) SCOTE, 7) NI-MWMOTE, 8) MWMOTE The experimental results are shown in Table 1. The experimental results show that the resampling performance of the present invention is better than other compared methods.

[0062] Table 1 Comparison of resampling performance results:

[0063] Please refer to Figure 2 , the second aspect of the present application provides an adaptive sampling system for fault diagnosis under data multi-class imbalance, including: A partitioning unit 201, configured to partition each category of samples according to the original dataset, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method for denoising preprocessing; A processing unit 202, configured to perform feature processing on the denoised dataset, use the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data to obtain a series of intrinsic mode functions, calculate the Shannon entropy for each modal function, and extract its eigenvalue to obtain a reconstructed feature dataset; An adjustment unit 203, configured to calculate the sample imbalance ratio and the intra-class and inter-class sample difference information of different datasets according to the analysis of data distribution characteristics, and dynamically adjust the key parameters in the sampling process; An acquisition unit 204, configured to obtain the global optimal solution in the multi-classification modeling accuracy model solution optimization by defining an objective function for minimizing the classification error and adopting the Newton-Raphson evolutionary mechanism, and the Newton-Raphson evolutionary mechanism includes the Newton-Raphson search and the trap avoidance mechanism; An evaluation unit 205, configured to use a feature recombination mechanism to convert the generated optimal solution samples into new synthetic samples of the minority class, and use a fitness evaluation mechanism to evaluate the synthetic new samples, so that the quality of the newly generated samples meets the requirements of the diagnostic model; A diagnosis unit 206, configured to perform small-sample multi-class imbalanced data fault diagnosis by combining the MAESTE method with multi-class LS-SVM.

[0064] Please refer to Figure 3 , this application also provides an adaptive sampling device for fault diagnosis under multi-class imbalance of data, including: A processor 301, a memory 302, an input / output unit 303, and a bus 304; The processor 301 is connected to the memory 302, the input / output unit 303, and the bus 304; The memory 302 stores a program, and the processor 301 calls the program to execute any of the above methods.

[0065] This application also relates to a computer-readable storage medium, on which a program is stored. It is characterized in that when the program runs on a computer, the computer is enabled to execute any of the above methods.

[0066] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, device, and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0067] It should be noted that the technology proposed in the present invention can be executed on a PC, a mobile terminal, or a similar computing device. Please refer to Figure 4 , Figure 4 is a block diagram of the hardware platform structure for classifying and diagnosing the health data of key components of high-end equipment involved in the present invention.

[0068] The platform includes a processor 102 and a data memory 104. Among them, the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA.

[0069] The above terminal may further include a transmission device 106 for communication functions and an input / output device 108.

[0070] Professional operation and maintenance technicians or managers in this field can understand that Figure 4 the structure shown is only schematic and does not limit the structure of the above terminal.

[0071] Among them, the memory 104 can be used to store computer programs, such as software programs and modules of application software, such as the computer program corresponding to the steam turbine data classification prediction in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implements the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as a magnetic storage device, a flash memory, or other non-volatile solid-state memories.

[0072] In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories can be connected to the terminal through a network. Examples of the above networks include, but are not limited to, the Internet, the internal network of a power plant, a local area network, a mobile communication network, and combinations thereof.

[0073] The transmission device 106 is used to receive or send data via a network. The above network includes a wireless network provided by a communication provider of the terminal.

[0074] In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus can communicate with the Internet.

[0075] In one instance, the transmission device 106 can be a wifi module, which is used to communicate with the Internet wirelessly.

[0076] This application also relates to a computer-readable storage medium, on which a program is stored. When the program runs on a computer, the computer is enabled to execute the method described in the first aspect and any one of the first aspects.

[0077] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0078] In several embodiments provided by this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of devices or units can be in electrical, mechanical, or other forms.

[0079] The unit described as a separate component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, that is, it may be located in one place or may be distributed over multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0080] In addition, each functional unit in various embodiments of the present application may be integrated in a processing unit, may exist separately as individual physical units, or two or more units may be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0081] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

Claims

1. An adaptive sampling method for fault diagnosis under multi-class data imbalance, characterized in that, At least include the following steps: Divide each category of samples according to the original dataset, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method for denoising preprocessing; Perform feature processing on the denoised dataset. Use the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data to obtain a series of intrinsic mode functions. Calculate the Shannon entropy for each modal function, and extract its eigenvalue to obtain the reconstructed feature dataset; According to the analysis of data distribution characteristics, calculate the sample imbalance ratio and the sample difference information between intra-class and inter-class of different datasets, and dynamically adjust the key parameters in the sampling process; By defining an objective function for minimizing classification error, use the Newton-Raphson evolutionary mechanism to obtain the global optimal solution in the optimization of the multi-classification modeling accuracy model. The Newton-Raphson evolutionary mechanism includes Newton-Raphson search and trap avoidance mechanisms; Use the feature recombination mechanism to transform the generated optimal solution samples into new synthetic samples of the minority class, and use the fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnostic model; Adopt the MAESTE method combined with multi-class LS-SVM for fault diagnosis of small-sample multi-class imbalanced data.

2. The adaptive sampling method for fault diagnosis under data multi-class imbalance according to claim 1, characterized in that Divide each category of samples according to the original dataset, calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method for denoising preprocessing, including: Divide the original dataset into majority-class samples and minority-class samples according to the number of samples in each category. Calculate the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and use the k-nearest neighbor method to perform denoising preprocessing on each category of samples.

3. The adaptive sampling method for fault diagnosis under multi-class data imbalance according to claim 1, wherein Calculate the Shannon entropy for each modal function, and extract its eigenvalue to obtain the reconstructed feature dataset, including: Adopt the ICEEMDAN-Shannon energy entropy method to extract features from the data after denoising preprocessing and reconstruct the dataset.

4. The adaptive sampling method for fault diagnosis under multi-class data imbalance according to claim 1, characterized in that, According to the analysis of data distribution characteristics, calculate the sample imbalance ratio and the sample difference information between intra-class and inter-class of different datasets, and dynamically adjust the key parameters in the sampling process, including: Dynamically and adaptively calculate the number of fault samples to be generated according to different data characteristics, and then calculate the total number of features to be solved in the solution process according to the total number of samples to be generated and the number of features.

5. The adaptive sampling method for fault diagnosis under multi-class data imbalance according to claim 2, characterized in that By defining an objective function for minimizing classification error, use the Newton-Raphson evolutionary mechanism to obtain the global optimal solution in the optimization of the multi-classification modeling accuracy model. The Newton-Raphson evolutionary mechanism includes Newton-Raphson search and trap avoidance mechanisms, including: Define the objective function in the solution process to obtain the minimum classification error, and use the Newton-Raphson search criterion to explore and solve for the global optimal solution.

6. The adaptive sampling method for fault diagnosis under multi-class data imbalance according to claim 1, wherein Use the feature recombination mechanism to transform the generated optimal solution samples into new synthetic samples of the minority class, and use the fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnostic model, including: Use the feature recombination mechanism to recombine the optimal solution to generate minority-class fault samples; Perform fitness evaluation on the generated samples to ensure that the quality of the newly generated samples meets the requirements; Adopt a fitness evaluation mechanism to evaluate by measuring the performance of the new samples in the classification model. If the fitness of the newly generated samples is better, update the samples. The expression of the fitness function is as follows: Among them, ACC is the classification error, G represents the error of the G-Mean index, F represents the F-Measure error, A represents the AUC error, and 𝛼1 and 𝛼2 are two weight parameters.

7. The adaptive sampling method for fault diagnosis under data multi-class imbalance according to claim 1, characterized in that Adopt the combination of MAESTE and multi-class LS-SVM for small-sample multi-class imbalanced data fault diagnosis, including: Merge the newly generated high-quality samples with the original sample set, and divide them into a training sample set and a test sample set. Adopt a one-versus-one strategy to train multiple LS-SVM models and classify through a voting mechanism; Use the test sample set as the input of the multi-class LS-SVM, calculate four evaluation indicators of accuracy, G-mean, F-measure, and AUC, and evaluate the MAESTE method through the four evaluation indicators; Adopt the combination of the MAESTE method and multi-class LS-SVM for small-sample multi-class imbalanced data fault diagnosis.

8. An adaptive sampling system for fault diagnosis under multi-class data imbalance, characterized in that, The system includes: A division unit for dividing each category of samples according to the original data set, calculating the Euclidean distances between each category of samples and their k nearest neighbors respectively to identify noise points, and using the k-nearest neighbor method for denoising preprocessing; A processing unit for performing feature processing on the denoised data set, using the ICEEMDAN modal decomposition method to perform multi-level decomposition on the vibration signal data, obtaining a series of intrinsic mode functions, calculating the Shannon entropy of each modal function, and extracting its eigenvalues to obtain the reconstructed feature data set; An adjustment unit for calculating the sample imbalance ratio and the intra-class and inter-class sample difference information of different data sets according to the analysis of the data distribution characteristics, and dynamically adjusting the key parameters in the sampling process; An acquisition unit for obtaining the global optimal solution in the multi-classification modeling accuracy model solution optimization by defining an objective function that minimizes the classification error and adopting the Newton-Raphson evolutionary mechanism, and the Newton-Raphson evolutionary mechanism includes a Newton-Raphson search and a trap avoidance mechanism; An evaluation unit for using a feature recombination mechanism to transform the generated optimal solution samples into new synthesized samples of the minority class, and adopting a fitness evaluation mechanism to evaluate the synthesized new samples so that the quality of the newly generated samples meets the requirements of the diagnosis model; A diagnosis unit for adopting the combination of the MAESTE method and multi-class LS-SVM for small-sample multi-class imbalanced data fault diagnosis.

9. Adaptive sampling device for fault diagnosis under multi-class data imbalance, characterized in that, The device includes: A processor, a storage, an input / output unit, and a bus; The processor is connected to the storage, the input / output unit, and the bus; The storage stores a program, and the processor calls the program to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The program is stored on the computer-readable storage medium, and when the program is executed on the computer, it executes the method according to any one of claims 1 to 7.

Images