Incremental equipment fault diagnosis method based on sample retention

Abstract

The present invention provides an incremental equipment fault diagnosis method based on sample retention, relating to the technical field of equipment fault diagnosis. The method comprises: selecting data representing the equipment's fault state and health state from sensor data to form an original data set; preprocessing the data set to form a complete sample set; establishing a fault diagnosis model based on a convolutional neural network; training the fault diagnosis model; inputting the sample set into the fault diagnosis model to filter out samples with correctly classified fault states; applying a sample retention method based on a genetic algorithm to the filtered sample set to form a sample subset; and retraining the fault diagnosis model using the sample subset and newly added samples. This method optimizes the model parameters by retraining the model using the newly added samples and the important sample subset based on the original model structure and parameters, thereby enabling the model to not only discriminate against new fault characteristics but also retain existing knowledge, thereby improving the adaptability of the fault diagnosis model.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault diagnosis, and in particular to an incremental equipment fault diagnosis method based on sample retention. Background Art

[0002] The increasing complexity of modern systems places higher demands on safety and stability. Therefore, timely identification, diagnosis, and rapid recovery of equipment system failures are essential to avoid further economic losses and personal safety incidents. Current fault diagnosis models rely on the strong assumption that training and test samples are identically distributed. However, as equipment continues to operate, new fault characteristics and types may emerge, making the original model no longer applicable. Therefore, timely updates to fault diagnosis models are necessary.

[0003] Currently, the update process of fault diagnosis models built for deep learning usually requires combining new and old samples, that is, using a complete sample set in a complete batch training mode. However, there are such problems: the equipment has accumulated a large number of historical samples over a long period of operation, and the storage cost is relatively high. If the model is retrained using a complete sample set, the time cost and computing cost are relatively high. In addition, if only new samples are used to adjust the model parameters, the model will easily forget the old samples. Too high an update frequency will generate unnecessary consumption, while too low an update frequency will make it difficult to ensure the performance of the model. Compared with the above-mentioned classic update methods, models with incremental learning capabilities have the following advantages: (1) they can continuously learn new knowledge from continuously generated new samples and can also retain most of the old knowledge; (2) they do not need to save all historical samples, only key samples, thus reducing storage space; (3) they make full use of historical training results to improve model training efficiency.

[0004] Currently, a considerable number of researchers are extending algorithmic functionality based on machine learning, deep learning, and other methods, enabling diagnostic models to acquire incremental learning capabilities. Because deep learning algorithms offer advantages over traditional machine learning algorithms in abstract feature representation, their application in fault diagnosis is gaining increasing attention. However, most methods for implementing incremental learning in fault diagnosis models based on deep learning involve expanding the model's structure and parameters. However, each model update requires the careful design of additional network structures. As incremental learning progresses, the redundancy of the model structure increases, leading to higher storage and computational costs. For example, the Chinese patent application "An Incremental Stacked Width Learning System with a Deep Structure" (Publication No. CN112508192A) incrementally updates the model by stacking multiple width learning systems. Although only newly added parameters are trained, the new model structure tends to be complex, and its feature mapping relationships are significantly limited by the fixed old parameters, which is not conducive to improving the model's fitting ability. Summary of the Invention

[0005] The purpose of the present invention is to provide an incremental equipment fault diagnosis method based on sample retention to address the above-mentioned deficiencies in the prior art, so as to solve the incremental learning problem of equipment fault diagnosis.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] The present invention provides an incremental device fault diagnosis method based on sample retention, the method comprising:

[0008] Select data representing various fault states and health states of the equipment from the sensor data that monitors the equipment status to form an original data set;

[0009] Performing preprocessing operations on the original data set to form a complete sample set. The preprocessing operations include normalizing the sensor data in the original data set, then cutting the normalized sensor data into segments of a preset length, and finally reassembling the segments into two-dimensional matrix data to form a complete sample set.

[0010] A fault diagnosis model for the equipment is established based on a convolutional neural network. The output of the fault diagnosis model is the probability that the input sample belongs to various fault states of the equipment.

[0011] Using the complete sample set to train the fault diagnosis model to form a trained fault diagnosis model;

[0012] Inputting the complete sample set into the trained fault diagnosis model to filter out samples with correct fault status classification from the complete sample set, and discarding samples with incorrect fault status classification in the complete sample set, so that the filtered samples with correct fault status classification constitute a filtered sample set;

[0013] For the screened sample set, a sample retention method based on a genetic algorithm is applied to select and retain specific samples, which form a sample subset. The statistical properties of the sample subset are used to characterize the statistical properties of the complete sample set.

[0014] The trained fault diagnosis model is retrained using the sample subset and the newly added samples to obtain a fault diagnosis model for equipment fault diagnosis, where the newly added samples are new samples different from the data samples in the original data set.

[0015] Optionally, during the process of training the fault diagnosis model, an Adam optimizer is used to optimize the parameters of the fault diagnosis model.

[0016] Optionally, a complete sample set is used to train a fault diagnosis model to form a trained fault diagnosis model, including:

[0017] a1) Initialize the parameters and weights of the fault diagnosis model;

[0018] b1) Randomly divide the complete sample set into batches, input the batches into the fault diagnosis model, and obtain the output value of each layer of the network;

[0019] c1) Based on the preset loss function and the current network model, calculate the error between the output value of each batch and the preset target value, and finally calculate the average error of all batches;

[0020] d1) Propagate the error backward from the output layer, calculate the gradient of the weight layer by layer, and update all network weights after completing all gradient calculations;

[0021] e1) Repeat steps b1) to d1) until a preset iteration stop condition is met, thereby forming a trained fault diagnosis model.

[0022] Optionally, the iteration stopping condition in step e1) is that the convergence iteration of the fault diagnosis model reaches a preset maximum number of iterations.

[0023] Optionally, a sample retention method based on a genetic algorithm is applied to the screened sample set to select and retain specific samples, and the specific samples constitute a sample subset, including:

[0024] a2) performing binary encoding on the index of the filtered sample set;

[0025] b2) Randomly initialize the population;

[0026] c2) Calculate the fitness of the current population;

[0027] d2) Using roulette wheel selection, two-point crossover, and multi-point mutation as selection operators, crossover operators, and mutation operators to generate new populations;

[0028] e2) Repeat steps c2) and d2) until a preset iteration stop condition is met;

[0029] f2) Decode the final population to obtain sample indexes, and construct sample subsets based on the corresponding samples.

[0030] Optionally, step c2) includes: decoding the population generated by the current iteration to obtain a corresponding current sample subset;

[0031] Input the filtered sample set and the current sample subset into the trained fault diagnosis model to obtain the corresponding logits vector;

[0032] Calculate the mean center of each logits vector and μ;

[0033] Calculating fitness Let F be the fitness of the current population.

[0034] The beneficial effects of the present invention include:

[0035] The present invention provides an incremental equipment fault diagnosis method based on sample retention, which includes: selecting data representing various fault states and health states of the equipment from sensor data for monitoring the equipment status to form an original data set; performing a preprocessing operation on the original data set to form a complete sample set, wherein the preprocessing operation includes normalizing the sensor data in the original data set, then cutting the normalized sensor data into segments with a preset length, and finally reassembling the segments into two-dimensional matrix data to form a complete sample set; establishing a fault diagnosis model for the equipment based on a convolutional neural network, wherein the output of the fault diagnosis model is the probability that the input sample belongs to various fault states of the equipment; using the complete sample set to train the fault diagnosis model to form a complete sample set. A trained fault diagnosis model is provided; a complete sample set is input into the trained fault diagnosis model to filter out samples with correct fault status classification from the complete sample set, and samples with incorrect fault status classification from the complete sample set are discarded, with the filtered samples with correct fault status classification constituting a filtered sample set; a sample retention method based on a genetic algorithm is applied to the filtered sample set to select and retain specific samples, which constitute a sample subset, and the statistical characteristics of the sample subset are used to characterize the statistical characteristics of the complete sample set; the trained fault diagnosis model is retrained using the sample subset and the newly added samples to obtain a fault diagnosis model for equipment fault diagnosis, wherein the newly added samples are new samples different from the data samples in the original dataset. This method uses the newly added samples and the important sample subset to jointly retrain the model based on the original model network structure and parameters, adjusting and optimizing the model parameters. This method not only enables the model to discriminate new fault characteristics and new fault types, but also retains most of the old knowledge, that is, it still maintains a relatively good memory level of historical samples, further improving the adaptability of the fault diagnosis model. The proposed method is applied to the field of equipment fault diagnosis and can continuously, adaptively, objectively and accurately reflect the fault type of the monitored equipment, providing a basis for guiding subsequent equipment fault management and maintenance work. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1A schematic diagram showing a flow chart of an incremental device fault diagnosis method based on sample retention provided by an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of cutting and converting sensor data provided by an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of chromosome encoding provided by an embodiment of the present invention;

[0040] Figure 4 Flowchart of the GA-based sample retention method provided in an embodiment of the present invention;

[0041] Figure 5A The experimental results of the diagnostic accuracy of the incremental equipment fault diagnosis method based on sample retention on the old category test set after completing one incremental learning;

[0042] Figure 5B The present invention provides an incremental device fault diagnosis method based on sample retention, which is provided in an embodiment of the present invention, and demonstrates the diagnostic accuracy experimental results on a new category test set after completing one incremental learning. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0044] In machine learning, historical samples significantly enhance model performance, but overly large sample sets can lead to high training and storage overhead. By selecting and retaining key samples from the complete sample set, and constructing a subset of important samples to replace the complete sample set, data storage costs and model retraining costs can be significantly reduced.

[0045] Figure 1 FIG. 4 shows a flow chart of an incremental device fault diagnosis method based on sample retention provided by an embodiment of the present invention, as shown in FIG. Figure 1 As shown, the method includes:

[0046] Step 101: Select data representing various fault states and health states of the equipment from sensor data for monitoring equipment status to form an original data set.

[0047] The status data is required to be continuously measured and recorded. Different types of industrial equipment may need to monitor different statuses, such as vibration, temperature, current, voltage and other signals.

[0048] Step 102: Preprocess the original data set to form a complete sample set.

[0049] The preprocessing operation includes normalizing the sensor data in the original dataset, cutting the normalized sensor data into segments of a preset length, and finally reassembling the segments into a two-dimensional matrix data to form a complete sample set. Specifically, first, the sensor data is normalized and the values ​​are limited to [0,1]. Then, as Figure 2 As shown in the figure, each sensor data point is cut into segments of length L. To increase the number of samples, overlapping sampling is used, meaning that there is a certain overlap between the previous and next samples. Finally, since most deep neural network models operate on multidimensional tensor data, the input data needs to be processed and the cut segments reassembled into a two-dimensional matrix data of H ╳ W to form a complete sample set, where H and W represent the number of rows and columns of the two-dimensional matrix, respectively, and L = H ╳ W.

[0050] Step 103: Establish a fault diagnosis model for the equipment based on the convolutional neural network.

[0051] The fault diagnosis model outputs the probability that an input sample belongs to various device fault states. Specifically, the fault diagnosis model is built based on a convolutional neural network (CNN). By stacking multiple convolutional layers and incorporating skip connections, a deep neural network model with excellent feature extraction capabilities is designed. Fully connected layers and softmax layers are used to implement multi-classification tasks. The input feature map size and number of channels are determined by the sample size (H╳W) and the number of data sources. The final output is a probability vector for each fault class for the input sample.

[0052] Step 104: Use the complete sample set to train the fault diagnosis model to form a trained fault diagnosis model.

[0053] All model training-related hyperparameters, such as the learning rate and maximum number of iterations, are pre-set. The fault diagnosis model is then trained using the complete sample set obtained in step 102. Optionally, during training, the Adam optimizer is used to optimize the model's parameters. This yields a high-performance fault diagnosis model that can be directly applied to equipment fault diagnosis.

[0054] Optionally, the complete sample set is used to train the fault diagnosis model to form a trained fault diagnosis model, specifically including:

[0055] a1) Initialize the parameters and weights of the fault diagnosis model;

[0056] b1) Randomly divide the complete sample set into batches, input the batches into the fault diagnosis model, and obtain the output value of each layer of the network;

[0057] c1) Based on the preset loss function and the current network model, calculate the error between the output value of each batch and the preset target value, and finally calculate the average error of all batches;

[0058] d1) Propagate the error backward from the output layer, calculate the gradient of the weight layer by layer, and update all network weights after completing all gradient calculations;

[0059] e1) Repeat steps b1) to d1) until a preset iteration stop condition is met, thereby forming a trained fault diagnosis model.

[0060] Optionally, the iteration stopping condition in step e1) is that the convergence iteration of the fault diagnosis model reaches a preset maximum number of iterations.

[0061] Step 105: Input the complete sample set into the trained fault diagnosis model to filter out samples with correct fault status classification from the complete sample set, and discard samples with incorrect fault status classification in the complete sample set. The filtered samples with correct fault status classification constitute the filtered sample set.

[0062] The previously obtained complete sample set is input into the fault diagnosis model to obtain the corresponding classification results. To prevent misclassified samples from causing unnecessary perturbations to subsequent incremental learning, the sample set needs to be screened. Only samples correctly predicted by the model enter the sample retention phase, while samples with incorrect predictions are discarded. This results in a correctly classified sample set.

[0063] Step 106: Apply a sample retention method based on a genetic algorithm to the screened sample set to select and retain specific samples, and the specific samples constitute a sample subset.

[0064] The statistical properties of the sample subset are used to characterize the statistical properties of the complete sample set.

[0065] Optionally, for the screened sample set, a sample retention method based on a genetic GA algorithm is applied to select and retain specific samples, and the specific samples constitute a sample subset, including (e.g. Figure 4 shown):

[0066] a2) Binary encode the index of the filtered sample set, specifically, chromosome encoding. The chromosome adopts binary encoding, and its length is equal to the size of the sample subset to be constructed. Each gene corresponds to a sample in the complete sample set. Figure 3 As shown in the figure, "11" represents a sample with index subscript 11 in the complete sample set, which is encoded as gene "1011" in binary, and the rest are similar.

[0067] b2) Randomly initialize the population;

[0068] c2) Calculate the fitness of the current population; specifically, decode the population generated by the current iteration to obtain the corresponding current sample subset; input the filtered sample set and the current sample subset obtained in step 105 into the trained fault diagnosis model to obtain the corresponding logits vector (that is, the output vector of the last fully connected layer); calculate the mean center of each logits vector and μ; calculate fitness Let F be the fitness of the current population;

[0069] d2) Using roulette wheel selection, two-point crossover, and multi-point mutation as selection operators, crossover operators, and mutation operators to generate new populations;

[0070] e2) Repeat steps c2) and d2) until a preset iteration stop condition is met;

[0071] f2) Decode the final population to obtain sample indexes, and construct sample subsets based on the corresponding samples.

[0072] Step 107: Retrain the trained fault diagnosis model using the sample subset and the newly added samples to obtain a fault diagnosis model for equipment fault diagnosis.

[0073] The newly added samples are new samples that are different from the data samples in the original dataset.

[0074] The present invention proposes an incremental equipment fault diagnosis method based on sample retention, establishes a fault diagnosis model based on a deep convolutional neural network (CNN), applies a sample retention method based on a genetic algorithm (GA) to construct a representative sample subset to replace the complete sample set, reduces storage resource requirements, and retrains the model using the sample subset and the new sample set on the basis of the original fault diagnosis model structure and parameters to achieve incremental learning of equipment fault diagnosis. On the basis of the original model network structure and parameters, the newly added samples and the important sample subset are jointly retrained to adjust and optimize the model parameters; so that it not only has the ability to discriminate new fault features and new fault types, but also can retain most of the old knowledge, that is, it still maintains a relatively good memory level for historical samples, further improving the adaptability of the fault diagnosis model.

[0075] To validate the performance of the proposed method, a case study and analysis was conducted using a bearing dataset from Case Western Reserve University (CWRU). The CWRU dataset includes three accelerometers collecting vibration data from different ends: the drive-end accelerometer (DE), the fan-end accelerometer (FE), and the basic acceleration data (BA). The dataset also includes four operating speeds of 1730, 1750, 1772, and 1797 rpm. The dataset also includes three bearing diameters of 0.007", 0.014", and 0.021". Furthermore, the dataset includes three fault conditions: inner race (IR), rolling element (Ball), and outer race (OR). The outer race fault condition also includes three measurement points: at 6 o'clock, directly in the load zone; at 3 o'clock, perpendicular to the load zone; and at 12 o'clock, opposite the load zone.

[0076] This paper uses state data from three bearings with diameters of 0.007", 0.014", and 0.021", operating at 1797 rpm. These data cover three types of faults: IR, Ball, and OR (load position 6:00). The bearing state data includes DE, FE, and BA data, sampled at a frequency of 12 kHz. The fault states in the dataset are classified into nine categories. The specific fault types are listed in Table 1, ranging from 0 to 8. The data volume for each fault type is shown in Table 2. The model structure parameters are shown in Table 3, where c represents the number of fault categories.

[0077] Table 1 Fault type numbers in the CWRU dataset

[0078]

[0079]

[0080] Table 2 Data volume of each fault type in the CWRU dataset

[0081]

[0082] Table 3 Network model parameters

[0083]

[0084]

[0085] The fault categories 0-9 are divided into two parts, the category number range 0-4 is the old category, and the number range 5-8 is the new category. The experiment starts with using the old category samples to train the fault diagnosis model, and then incrementally learns the new category samples through the sample retention method. In order to show the advantages of this application, the NME-based sample retention method and the random sampling-based sample retention method commonly used in the incremental learning field are used for comparative analysis. The experimental results of incremental equipment fault diagnosis based on sample retention are shown in Figure 2. Figure 5A and Figure 5B As shown, Figure 5A and Figure 5B They are the diagnostic accuracy of the fault diagnosis model on the old category and new category test sets after completing one incremental learning. Figure 5B Regardless of the sample retention method used, the diagnostic accuracy of the model on the new category test set is higher than 99.80%, indicating that the fault diagnosis model effectively learns the fault characteristics of the new category samples in each incremental learning.

[0086] 1) Analysis of the impact of sample retention methods

[0087] Figure 5A The model diagnostic accuracy of the GA-based sample retention method under different sample retention numbers is 76.97%, 95.14%, 98.17%, and 99.42%, the NME-based sample retention method is 71.37%, 89.09%, 91.09%, and 93.54%, and the random sampling-based sample retention method is 74.98%, 91.11%, 96.53%, and 98.99%. Obviously, the GA-based sample retention method has always maintained the best performance among the three, and it also reaches accuracy saturation (close to 100%) the fastest. Secondly, in this case, the performance of the NME-based sample retention method is worse than that of the random sampling-based sample retention method, indicating that it is very necessary to study more effective sample retention methods.

[0088] 2) Analysis of the impact of sample retention quantity

[0089] Figure 5AThe results show that when the number of retained samples increases from 5 to 30, the diagnostic accuracy of the models using the NME-based, GA-based, and random sampling-based methods for incremental learning increases from 71.37% to 93.54%, 76.97% to 99.42%, and 74.98% to 98.99%, respectively. This indicates that the more samples retained, the higher the model's diagnostic accuracy. When the number of retained samples for each old category reaches 20 or 30, the fault diagnosis model's discriminative ability for that category is already quite good. Clearly, in this case, the important sample subset can almost fully characterize the characteristics of the complete sample set. Therefore, for the CWRU dataset fault diagnosis case, retaining all samples is not necessary. Simply selecting and retaining samples to construct an important sample subset can significantly improve the incremental learning of the fault diagnosis model.

[0090] 3) Training time

[0091] The model training time corresponding to the above experimental method is shown in Table 4. Full batch training refers to training the model with new and old samples together. Since the test accuracy of this method on 9 fault categories is 99.00% and above, the diagnostic accuracy of the full batch training method is not Figure 5A and Figure 5B Table 4 shows that when the number of retained samples is 30, the model accuracy is relatively close to that of full batch training, but the training time for the former is 25.45s, while the latter is 46.67s. This shows that the sample retention method can significantly reduce the model update time, indicating that the incremental learning method based on sample retention has an efficiency advantage in model training.

[0092] Table 4 Model training time

[0093]

[0094]

[0095] This invention establishes a fault diagnosis model based on deep learning, boasting excellent nonlinear feature representation and fitting capabilities. It can extract key features from a large number of samples and accurately identify fault categories. This application selects key samples from historical samples for retention, constructs a subset of important samples, and uses the statistical properties of this subset to characterize the statistical properties of the complete historical sample set, thereby reducing data memory consumption. Because fault diagnosis models struggle to effectively discriminate against samples they have never learned, this application addresses the problem of incremental learning in fault diagnosis models. Based on the existing model network structure and parameters, the model is retrained using newly added samples and a subset of important samples, adjusting and optimizing the model parameters. This approach not only enables the model to discriminate against new fault features and types, but also retains much of the previous knowledge, maintaining a relatively good memory of historical samples, further improving the adaptability of the fault diagnosis model. Applied to the field of equipment fault diagnosis, this proposed method can consistently, adaptively, objectively, and accurately reflect the fault types of monitored equipment, providing a basis for guiding subsequent equipment fault management and maintenance.

[0096] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable ordinary technicians in this field to understand the content of the present invention and implement it. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. An incremental equipment fault diagnosis method based on sample retention, characterized in that: The method comprises: Select data representing various fault states and health states of the equipment from the sensor data that monitors the equipment status to form an original data set; performing a preprocessing operation on the original data set to form a complete sample set, the preprocessing operation comprising normalizing the sensor data in the original data set, then cutting the normalized sensor data into segments of a preset length, and finally reassembling the segments into two-dimensional matrix data to form a complete sample set; Establishing a fault diagnosis model for the device based on a convolutional neural network, wherein the output of the fault diagnosis model is the probability that the input sample belongs to multiple fault states of the device; Using the complete sample set to train the fault diagnosis model to form a trained fault diagnosis model; Inputting the complete sample set into the trained fault diagnosis model to filter out samples with correct fault status classification from the complete sample set, and discarding samples with incorrect fault status classification from the complete sample set, so that the filtered samples with correct fault status classification constitute a filtered sample set; For the screened sample set, applying a sample retention method based on a genetic algorithm to select and retain specific samples, forming a sample subset from the specific samples, and using the statistical properties of the sample subset to characterize the statistical properties of the complete sample set; The trained fault diagnosis model is retrained using the sample subset and newly added samples to obtain a fault diagnosis model for equipment fault diagnosis, wherein the newly added samples are new samples different from the data samples in the original data set.

2. The incremental device fault diagnosis method based on sample retention according to claim 1, characterized in that: During the process of training the fault diagnosis model, an Adam optimizer is used to optimize the parameters of the fault diagnosis model.

3. The incremental equipment fault diagnosis method based on sample retention according to claim 1, characterized in that: The using the complete sample set to train the fault diagnosis model to form a trained fault diagnosis model includes: a1) initializing the parameters and weights of the fault diagnosis model; b1) randomly dividing the complete sample set into batches, inputting the batches into the fault diagnosis model, and obtaining the output value of each layer of the network; c1) Based on the preset loss function and the current network model, calculate the error between the output value of each batch and the preset target value, and finally calculate the average error of all batches; d1) Propagate the error backward from the output layer, calculate the gradient of the weight layer by layer, and update all network weights after completing all gradient calculations; e1) Repeat steps b1) to d1) until a preset iteration stop condition is met, thereby forming a trained fault diagnosis model.

4. The incremental equipment fault diagnosis method based on sample retention according to claim 3 is characterized in that: The iteration stopping condition in step e1) is that the convergence iteration of the fault diagnosis model reaches a preset maximum number of iterations.

5. The incremental equipment fault diagnosis method based on sample retention according to claim 1, characterized in that: The method of applying a sample retention method based on a genetic algorithm to the screened sample set to select and retain specific samples, and forming a sample subset from the specific samples, includes: a2) performing binary encoding on the filtered sample set index; b2) Randomly initialize the population; c2) Calculate the fitness of the current population; d2) Using roulette wheel selection, two-point crossover, and multi-point mutation as selection operators, crossover operators, and mutation operators to generate new populations; e2) Repeat steps c2) and d2) until a preset iteration stop condition is met; f2) Decode the final population to obtain sample indexes, and construct sample subsets based on the corresponding samples.

6. The incremental equipment fault diagnosis method based on sample retention according to claim 5, characterized in that: Step c2) comprises: Decode the population generated by the current iteration to obtain the corresponding current sample subset; Inputting the filtered sample set and the current sample subset into the trained fault diagnosis model to obtain corresponding logits vectors; Calculate the mean center of each logits vector and μ; Calculating fitness Let F be the fitness of the current population.

Images