A missing sample imputation generation method and device under unknown fault degree and electronic equipment

Abstract

Embodiments of the present application provide a missing sample imputation generation method and device under unknown fault degree and electronic equipment. The method comprises: extracting degradation trend conditional variables capable of representing equipment fault degree from each fault sample under existing partial known fault degree, and constructing a training sample set using the degradation trend conditional variables of each fault sample; constructing a CRGAN-MDTR network model comprising a generator network, and training the CRGAN-MDTR network model comprising the generator network using the training sample set to obtain a trained generator network; obtaining degradation trend conditional variables under a target fault degree, and inputting the degradation trend conditional variables under the target fault degree into the trained generator network to generate fault samples whose fault degree corresponds to the degradation trend conditional variables under the target fault degree, thereby achieving missing fault sample imputation generation under unknown fault degree. The present application can quickly output defect fault samples, thereby improving the efficiency of obtaining fault samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electromechanical product fault sample generation, and in particular to a missing sample interpolation generation method and device under unknown fault degree and electronic equipment. BACKGROUND

[0002] Electromechanical products and components have dynamic degradation characteristics. When the degradation degree exceeds the acceptable limit, the product cannot work normally, that is, it shows failure externally. Since the acceptable limit that affects the normal operation of the product is usually not fixed, the fault degree shown externally when the fault occurs is also uncertain, and there are obvious differences between fault samples under different fault degrees. Considering the uncertainty of fault degree, sample difference under different fault degrees and other factors, in order to meet the independent and identically distributed basic requirements required by the fault diagnosis model of machine learning type data driven, fault samples under various fault degrees need to be included in the training data set. However, when the actual fault sample quantity is small, especially when the sample is missing under part unknown fault degree, it is also difficult to obtain fault samples under different fault degrees through experiments. The fault injection method has low coverage of the degradation process, and it is difficult to perform fault injection experiments one by one for various possible fault degrees. The acceleration test method covers all fault degrees in the degradation process, but the time and economic cost required is very high. SUMMARY

[0003] The present application provides a missing sample interpolation generation method and device under unknown fault degree and electronic equipment, which is used to solve how to obtain fault samples under different fault degrees in the product degradation process at low cost.

[0004] In one aspect of the present application, a missing sample interpolation generation method under unknown fault degree is provided, comprising: extracting degradation trend conditional variables capable of representing equipment fault degree from each fault sample under part known fault degree, and constructing a training sample set by using the degradation trend conditional variables of each fault sample; constructing a CRGAN-MDTR network model containing a generator network, and training the CRGAN-MDTR network model containing the generator network by using the training sample set to obtain a trained generator network; obtaining degradation trend conditional variables under a target fault degree, and generating fault samples with degradation trend conditional variables corresponding to the target fault degree by inputting the degradation trend conditional variables under the target fault degree into the trained generator network, thereby realizing the interpolation generation of missing fault samples under unknown fault degree; wherein the CRGAN-MDTR refers to a conditional time series generative adversarial network model with maximum degradation trend retention rate.

[0005] Preferably, the extracting degradation trend conditional variables capable of representing the equipment failure degree from each failure sample under the existing partially known failure degree, and constructing a training sample set by using the degradation trend conditional variables of each failure sample comprises: extracting a plurality of features capable of representing the equipment failure degree from each failure sample, and splicing the plurality of features capable of representing the equipment failure degree to form a degradation trend conditional variable corresponding to the failure sample; splicing each failure sample and the degradation trend conditional variable extracted from the failure sample to form a failure sample pair with conditional information, and arranging all failure sample pairs in order of failure degree from small to large; and arranging all failure sample pairs arranged in order of failure degree from small to large as the training sample set.

[0006] Preferably, the method further comprises: sampling the same number of noise vectors as the failure sample pairs in the training sample set from a predetermined prior noise distribution, and splicing the noise vectors into a matrix by rows; copying and splicing the degradation trend conditional variables of the failure sample pairs into the noise vector matrix, so that the conditional variables attached to the failure samples and the noise vectors are the same, and taking the noise vector matrix spliced with the degradation trend conditional variables of the failure sample pairs as training noise vectors.

[0007] Preferably, the training the CRGAN-MDTR network model containing the generator network by using the training sample set to obtain a trained generator network comprises: inputting the training sample set and the training noise vectors into the CRGAN-MDTR network model containing the generator network to train the CRGAN-MDTR network model, extracting a trained generator network from the trained CRGAN-MDTR network model.

[0008] Preferably, the CRGAN-MDTR network model further comprises a discriminator network and a degradation trend reconstruction network, wherein the training of the CRGAN-MDTR network model by inputting the training sample set and the training noise vector into the CRGAN-MDTR network model comprising the generator network comprises: inputting the training noise vector into the generator network of the CRGAN-MDTR network model to obtain a generated sample, and inputting the generated sample into the degradation trend reconstruction network of the CRGAN-MDTR network model to obtain a reconstructed degradation trend condition variable; calculating an interpolation sample according to a fault sample pair in the training sample set and the generated sample, and calculating a discriminator loss by using the fault sample pair in the training sample set, the generated sample and the interpolation sample; calculating a generator loss according to a discrimination result of the generated sample by the discriminator; calculating a conditional mutual information maximization loss of the generator network and the degradation trend reconstruction network according to a reconstruction result of the degradation trend condition variable contained in the generated sample by the degradation trend reconstruction network; and sequentially performing optimization update on the discriminator network, the generator network and the degradation trend reconstruction network based on the discriminator loss, the generator loss and the conditional mutual information maximization loss until the loss converges, thereby completing the training of the CRGAN-MDTR network model and obtaining the trained CRGAN-MDTR network model.

[0009] Preferably, the generating of the fault sample with the degradation trend condition variable corresponding to the target fault degree by inputting the degradation trend condition variable under the target fault degree into the trained generator network comprises: sampling a noise vector from a Gaussian distribution, and obtaining a target noise vector by splicing the degradation trend condition variable under the target fault degree on the noise vector; and inputting the target noise vector and the degradation trend condition variable under the target fault degree into the trained generator network to obtain the fault sample with the degradation trend condition variable corresponding to the target fault degree.

[0010] Preferably, after obtaining the fault sample with the degradation trend condition variable corresponding to the target fault degree, the method further comprises: calculating the degradation trend condition variable under an unknown fault degree by interpolation between the degradation trend condition variables extracted from the existing fault samples, thereby completing the interpolation generation of the missing fault sample under the unknown fault degree.

[0011] In another aspect of the present application, a missing sample imputation generation device under unknown fault degree is provided, comprising: a training sample set construction module configured to extract a degradation trend conditional variable capable of representing the fault degree of the equipment from each fault sample under a known fault degree, and to construct a training sample set using the degradation trend conditional variable of each fault sample; a model construction and training module configured to construct a CRGAN-MDTR network model comprising a generator network, and to train the CRGAN-MDTR network model comprising the generator network using the training sample set to obtain a trained generator network; and a sample imputation generation module configured to obtain a degradation trend conditional variable under a target fault degree, and to input the degradation trend conditional variable under the target fault degree into the trained generator network to generate a fault sample whose degradation trend conditional variable corresponds to the target fault degree, thereby realizing the imputation generation of a missing fault sample under an unknown fault degree; wherein the CRGAN-MDTR refers to a conditional time series generative adversarial network model with maximum degradation trend retention rate.

[0012] Preferably, the training sample set construction module is specifically configured to extract a plurality of features capable of representing the fault degree of the equipment from each fault sample, and to splice the plurality of features capable of representing the fault degree of the equipment to form a degradation trend conditional variable corresponding to the fault sample; to splice each fault sample and the degradation trend conditional variable extracted from the fault sample to form a fault sample pair with conditional information, and to arrange all fault sample pairs in order of fault degree from small to large; and to arrange all fault sample pairs in order of fault degree from small to large as a training sample set.

[0013] In another aspect of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the missing sample imputation generation method under unknown fault degree are implemented.

[0014] The missing sample imputation generation method, device and electronic device under unknown fault degree provided by the embodiments of the present application use the pre-trained generator network to process the input current noise vector and the pre-set conditional variable of the expected generated sample fault degree, thereby outputting a defect fault sample matched with the conditional variable, so that the fault samples under different fault degrees in the product degradation process can be obtained at low cost. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and other objects, features and advantages of the present application will become readily apparent from the detailed description that follows, read in conjunction with the accompanying drawings. In the drawings, which represent several embodiments of the present application, the following is shown:

[0016] Figure 1 A flowchart of a missing sample imputation generation method under unknown fault degree provided by an embodiment of the present application;

[0017] Figure 2 A schematic diagram of a missing sample imputation generation device under unknown fault degree provided by an embodiment of the present application;

[0018] Figure 3 A schematic diagram of a missing sample imputation generation method under unknown fault degree based on CRGAN-MDTR provided by an embodiment of the present application;

[0019] Figure 4 A schematic diagram of a missing fault sample imputation generation provided by an embodiment of the present application;

[0020] Figure 5 A schematic diagram of a conditional time series deep generative adversarial network module structure provided by an embodiment of the present application;

[0021] Figure 6 A schematic diagram of a sample generation conditional mutual information maximization module structure provided by an embodiment of the present application;

[0022] Figure 7 A schematic diagram of bearing data anchor point setting and masking processing provided by an embodiment of the present application;

[0023] Figure 8 A schematic diagram of a CRGAN-MDTR model structure and hyperparameters provided by an embodiment of the present application;

[0024] Figure 9 A schematic diagram of an anchor point fault sample generation result provided by an embodiment of the present application;

[0025] Figure 10 A schematic diagram of a full life degradation process fault sample generation result provided by an embodiment of the present application;

[0026] Figure 11 A schematic diagram of a full life degradation process fault sample generation spectrum provided by an embodiment of the present application;

[0027] Figure 12 A schematic diagram of an SVM test diagnosis accuracy result under different n value conditions provided by an embodiment of the present application. DETAILED DESCRIPTION

[0028] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and so that the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0029] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have meanings consistent with meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined.

[0030] The deep generative model represented by the GAN is a kind of machine learning model that has rapidly developed in the field of artificial intelligence in recent years, belongs to the category of generative model, and has become one of the popular directions of research in the field of artificial intelligence. The present application proposes a conditional recurrent GAN with maximum degenerate trend retention (CRGAN-MDTR) model based on degenerate trend retention rate maximization. The model is based on the CGAN architecture. In terms of training structure, a time series LSTM unit is introduced. In terms of optimization objective, the Wasserstein loss and gradient penalty are introduced to strengthen the learning ability of CGAN for time series data and stabilize the training process. With the help of variational inference, the generated fault sample fault degree is indirectly constrained to match the fault degree input as the conditional control information, and the controllable generation of fault samples under a specific fault degree is realized.

[0031] Figure 1 The flowchart of the method for generating missing samples under unknown fault degree provided by the embodiments of the present application is shown in FIG. 1. Figure 1As shown, it comprises: step S1: extracting degradation trend conditional variables capable of representing the equipment failure degree from each failure sample under the existing partial known failure degree, and constructing a training sample set using the degradation trend conditional variables of each failure sample; step S2: constructing a CRGAN-MDTR network model containing a generator network, and training the CRGAN-MDTR network model containing the generator network using the training sample set to obtain a trained generator network; step S3: obtaining the degradation trend conditional variables under the target failure degree, and generating a failure sample whose failure degree corresponds to the degradation trend conditional variables under the target failure degree by inputting the degradation trend conditional variables under the target failure degree into the trained generator network, thereby realizing the interpolation generation of the missing failure sample under the unknown failure degree.

[0032] In the embodiment of the present application, the extracting degradation trend conditional variables capable of representing the equipment failure degree from each failure sample under the existing partial known failure degree, and constructing a training sample set using the degradation trend conditional variables of each failure sample comprises: extracting a plurality of features capable of representing the equipment failure degree from each failure sample, and splicing the plurality of features capable of representing the equipment failure degree to form a degradation trend conditional variable corresponding to the failure sample; splicing each failure sample and the degradation trend conditional variable extracted from the failure sample to form a failure sample pair with conditional information, and arranging all failure sample pairs in order of failure degree from small to large; and arranging all failure sample pairs arranged in order of failure degree from small to large as a training sample set.

[0033] In the embodiment of the present application, it further comprises: sampling the same number of noise vectors as the failure sample pairs in the training sample set from a predetermined prior noise distribution, and splicing the noise vectors into a matrix by rows; copying and splicing the degradation trend conditional variables of the failure sample pairs into the noise vector matrix, so that the conditional variables attached to the failure sample and the noise vector are the same, and taking the noise vector matrix spliced with the degradation trend conditional variables of the failure sample pairs as a training noise vector.

[0034] Further, the training the CRGAN-MDTR network model containing the generator network using the training sample set to obtain a trained generator network comprises: training the CRGAN-MDTR network model containing the generator network by inputting the training sample set and the training noise vector into the CRGAN-MDTR network model containing the generator network, obtaining a trained CRGAN-MDTR network model, and extracting a trained generator network from the trained CRGAN-MDTR network model.

[0035] More specifically, the CRGAN-MDTR network model further includes a discriminator network and a degradation trend reconstruction network, wherein the training by inputting the training sample set and the training noise vector into the CRGAN-MDTR network model including the generator network to obtain the trained CRGAN-MDTR network model includes: inputting the training noise vector into the generator network of the CRGAN-MDTR network model to obtain generated samples, and inputting the generated samples into the degradation trend reconstruction network of the CRGAN-MDTR network model to obtain reconstructed degradation trend condition variables; calculating the fault sample pairs in the training sample set and the generated samples according to the fault sample pairs in the training sample set. The interpolation samples are calculated, and the discriminator loss is calculated using the fault sample pairs in the training sample set, the generated samples and the interpolation samples; the generator loss is calculated according to the discrimination result of the discriminator on the generated samples; the conditional mutual information maximization loss of the generator network and the degradation trend reconstruction network is calculated according to the reconstruction result of the degradation trend conditional variables contained in the generated samples by the degradation trend reconstruction network; based on the discriminator loss, the generator loss and the conditional mutual information maximization loss, the discriminator network, the generator network and the degradation trend reconstruction network are optimized and updated in sequence until the loss converges, thereby completing the CRGAN-MDTR network model training and obtaining a trained CRGAN-MDTR network model.

[0036] In an embodiment of the present invention, the step of generating a fault sample whose fault degree corresponds to the degradation trend condition variable at the target fault degree by inputting the degradation trend condition variable at the target fault degree into the trained generator network includes: sampling a noise vector from a Gaussian distribution, and obtaining a target noise vector by splicing the degradation trend condition variable at the target fault degree onto the noise vector; and inputting the target noise vector and the degradation trend condition variable at the target fault degree into the trained generator network to obtain a fault sample whose fault degree corresponds to the degradation trend condition variable at the target fault degree.

[0037] In an embodiment of the present invention, after obtaining a fault sample whose fault degree corresponds to the degradation trend condition variable under the target fault degree, the method further includes: calculating the degradation trend condition variable under the unknown fault degree between the degradation trend condition variables extracted from the existing fault samples by interpolation, thereby completing the interpolation generation of missing fault samples under the unknown fault degree.

[0038] The present invention can quickly output defect fault samples, thereby improving the efficiency of obtaining fault samples.

[0039] Figure 2 Schematic diagram of a missing sample interpolation generation device under unknown fault degree provided by an embodiment of the present invention, such asFigure 2 As shown, comprising: a training sample set construction module, configured to extract degradation trend condition variables capable of representing the fault degree of the equipment from each of the existing partial known fault degree fault samples, and construct a training sample set by using the degradation trend condition variables of each fault sample; a model construction and training module, configured to construct a CRGAN-MDTR network model containing a generator network, and train the CRGAN-MDTR network model containing the generator network by using the training sample set, to obtain a trained generator network; and a sample interpolation generation module, configured to obtain degradation trend condition variables under a target fault degree, and generate a fault sample with degradation trend condition variables corresponding to the target fault degree by inputting the degradation trend condition variables under the target fault degree into the trained generator network, to realize interpolation generation of missing fault samples under unknown fault degrees.

[0040] Further, the training sample set construction module is specifically configured to extract a plurality of features capable of representing the fault degree of the equipment from each fault sample, and splice the plurality of features capable of representing the fault degree of the equipment to form a degradation trend condition variable corresponding to the fault sample; splice each fault sample and the degradation trend condition variable extracted from the fault sample to form a fault sample pair with condition information, and arrange all fault sample pairs in order of fault degree from small to large; and arrange all fault sample pairs in order of fault degree from small to large as a training sample set.

[0041] The present application can quickly output defect fault samples, and improves the efficiency of obtaining fault samples.

[0042] In another aspect of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the method for interpolating and generating missing samples under unknown fault degrees are implemented.

[0043] The present application proposes a conditional time sequence generative adversarial network CRGAN-MDTR model based on maximum degradation trend retention rate, for realizing interpolation generation of missing fault samples under unknown fault degrees. Figure 3 As shown, the specific process is as follows:

[0044] Step one: data preparation phase

[0045] The data preparation phase mainly completes extraction of known fault sample degradation trend condition variables, construction of a training sample set composed of fault samples and condition variables, and sampling of noise vectors.

[0046] In the degradation trend condition variable extraction stage, the fault samples under the known fault degree are sent into the degradation trend condition variable extraction module, k features representing the fault degree of the device are extracted from each fault sample, and the degradation trend condition variable corresponding to the fault sample is spliced. The training sample set construction is to splice the original fault sample and the degradation trend condition variable extracted from the sample to form a fault sample pair with condition information, and arrange all sample pairs in order from small to large fault degree. The noise vector sampling is to sample the same number of noise vectors as the fault sample pair from the predetermined prior noise distribution (such as Gaussian distribution, uniform distribution, etc.), and splice them into a matrix; then the degradation trend condition variable of the real fault sample pair is copied and spliced into the noise vector matrix, so that the real sample and the noise vector have the same condition variable.

[0047] where the noise vector z follows a Gaussian distribution P z or a uniform distribution.

[0048] Step two: model training stage

[0049] In the model training stage, the fault sample set and the noise sample set generated in the data preparation stage are first used to perform forward propagation, and the generated samples, the interpolated samples and the discriminator discrimination results are calculated. The Wasserstein loss of the discriminator is calculated according to the discrimination results of the real sample and the discrimination results of the generated sample, and the gradient penalty loss of the discriminator is calculated according to the interpolated sample. The above three loss (i.e. two Wasserstein losses and one gradient penalty loss) constitutes the discriminator loss of the conditional time series deep generative adversarial network module, corresponding to Figure 3 L D in the "model training" in the middle; the Wasserstein loss of the generator is calculated according to the discrimination results of the generated sample, corresponding to Figure 3 L G in the "model training" in the middle; the conditional mutual information maximization loss of the generator and the Q network is calculated according to the reconstruction result of the degradation trend condition variable contained in the generated sample by the degradation trend reconstruction network Q, corresponding to Figure 3 L DE in the "model training" in the middle. Based on the above loss terms, the discriminator, the generator and the Q network are sequentially optimized and updated until the loss converges, the model training is completed, and the trained generator network G * is obtained.

[0050] Step three: sample generation stage under unknown fault degree

[0051] The noise vector obtained is first sampled in this stage, and the degradation trend condition variable under the target fault degree is spliced on the noise vector. The noise vector and the degradation trend condition variable under the target fault degree are jointly input into the trained generator network to obtain the generator network output. The output is the fault sample corresponding to the input degradation trend condition variable under the fault degree. By interpolation, the degradation trend condition variable under the unknown fault degree is calculated between the degradation trend condition variables extracted from the existing fault samples, that is, the missing fault sample interpolation generation under the unknown fault degree is completed. The present application is a kind of interpolation generation technology based on continuous degradation trend fault, therefore the degradation trend condition variable is linearly changed with the deepening of fault degree. As shown in Figure 4 The ordinate is the degradation trend condition variable, and the abscissa is the fault degree.

[0052] In the sample generation part under unknown fault degree, the noise vector is sampled from Gaussian distribution, and the sampled noise vector satisfies the same shape as the real sample vector. Figure 3 In the sample generation part under unknown fault degree, the noise vector is sampled from Gaussian distribution, and the sampled noise vector satisfies the same shape as the real sample vector. g .

[0053] The missing fault sample interpolation generation method under unknown fault degree proposed by the present application can be applied under the following prerequisite conditions: the degradation process of the product is continuous and has no mutation phenomenon; the influence of local fluctuation phenomenon in the degradation process on the fault sample is much smaller than the influence of the overall degradation trend.

[0054] The CRGAN-MDTR model in the embodiment of the application is composed of three parts, which are a degradation trend conditional variable extraction module, a conditional time series deep generative adversarial network module and a generated sample conditional mutual information maximization module. The degradation trend conditional variable extraction module extracts a degradation trend conditional variable (Degradation Index, DI) capable of representing the equipment failure degree from the existing sample with a failure degree, as the conditional control information input to the generated model in the training process. The time series deep generative adversarial network module is based on the CGAN architecture, introduces a time series LSTM unit in the training structure, and introduces a Wasserstein loss and a gradient penalty in the optimization objective, to strengthen the learning ability of CGAN for time series data and stabilize the training process. The generated sample conditional mutual information maximization module combines the ideas of supervised and unsupervised conditional control sample generation, introduces a degradation trend reconstruction network in parallel based on the discriminator with sample and conditional information as double inputs, estimates the mutual information between the DI of the generated sample and the real DI input to the generator by means of variational inference, indirectly constrains the matching between the failure degree of the generated failure sample and the failure degree input as the conditional control information, and realizes the controllable generation of the failure sample under a specific failure degree.

[0055] The three modules are described below:

[0056] (1) Degradation trend conditional variable extraction module

[0057] The degradation trend conditional variable extraction module is responsible for extracting a DI capable of representing the equipment failure degree from the existing sample with a known failure degree. Common DI extraction methods include a physical feature-based extraction method and a data-driven model-based extraction method.

[0058] Taking a vibration signal as an example, the physical feature-based DI extraction method includes calculating the root mean square (RMS), peak factor, kurtosis, skewness and the like of the vibration signal sample.

[0059] For a vibration signal sample x=(x1, x2, …, x n ), the calculation method of RMS is as follows:

[0060]

[0061] The peak factor represents the ratio of the amplitude peak value of a signal sample to the effective value of the sample, and the calculation method is as follows:

[0062]

[0063] The calculation method of kurtosis is as follows:

[0064]

[0065] Where μ and σ represent the mean and standard deviation of the samples respectively, and n is the number of samples.

[0066] The calculation method of skewness is as follows:

[0067]

[0068] In practical applications, it is necessary to select appropriate physical features as DI based on the characteristics of the object and data.

[0069] Generally speaking, not all of these physical characteristics of an object can well reflect the degradation trend of the object ( Figure 4 Therefore, for different objects, using the aforementioned DI extraction method for physical features, we select a feature that most clearly represents the degradation trend as the degradation trend conditional variable. For the bearing vibration signal in this case, RMS is the most obvious feature representing the degradation trend and is therefore selected as the degradation trend conditional variable.

[0070] The DI extraction method based on a data-driven model uses healthy samples to construct a healthy baseline. The deviation between all samples and the healthy baseline is then calculated using a distance metric to represent the DI, which measures the degree of fault in the samples. Common distance metrics include Euclidean distance and Mahalanobis distance.

[0071] For a sample set X={x1,x2,…,x n After the degradation trend condition variable extraction module calculates the DI for each sample, each original sample is combined with its own DI into a sample pair with condition variables to form a training data set. where c i is a conditional variable vector composed of k DI values ​​extracted by different methods, c i =(DI1,DI2,…,DI k ).

[0072] (2) Conditional Temporal Deep Generative Adversarial Network Module

[0073] The degradation process of the product has typical time sequence characteristics, and there is a significant time sequence dependent relationship between the fault samples at different fault levels during the degradation process. A simple fully connected network can only learn the fault samples at different fault levels in a discrete manner and cannot mine the time sequence dependent relationship between them. LSTM is a representative module in time sequence neural networks and can effectively capture the long-distance time sequence dependent relationship between data and better model the samples at different fault levels in the degradation process which have typical time sequence characteristics. Therefore, in order to more effectively learn the degradation information contained in the fault samples at different fault levels, the LSTM unit is introduced into the CGAN architecture, and the LSTM unit with time sequence characteristic modeling and mining capability is used to replace the fully connected unit in the basic CGAN architecture. The module structure is as shown in Figure 5 .

[0074] Unlike the basic CGAN, the time sequence deep generative adversarial network with the introduction of the LSTM unit arranges the samples in each batch in the order of fault level from small to large during training. This is reflected in two aspects. On the one hand, when inputting the noise vectors in a batch into the generator one by one, the noise vectors are spliced with the degradation trend condition variable in the order from small to large. In this way, the fault level of the fault samples output by the generator should also show the order from small to large. On the other hand, when inputting the real fault samples into the discriminator, the fault level is also arranged in the order from small to large. The significance of introducing time sequence information into the generative adversarial network is that the discriminator network needs to not only judge whether the input sample is a real sample and whether the input sample matches the condition variable, but also judge whether the fault level between the current input sample and the previous and subsequent samples meets the order relationship, so as to promote the network to learn effective degradation information from the samples at different fault levels. In addition, the time sequence deep generative adversarial network module in the present application replaces the JS divergence loss in the basic CGAN with the Wasserstein distance loss, and introduces gradient penalty at the same time, so that the training is more stable and the quality of the generated samples is higher.

[0075] For a sample pair (x, c) composed of a real sample and a corresponding condition variable, a noise vector is sampled from random noise, and the condition variable is spliced to form a noise sample pair (z, c) with condition information. The noise sample pair (z, c) is input into the generator to obtain a generated sample x g =G(z, c) under the control of the condition variable c. The linear interpolation is performed on the real sample x and the generated sample x g to obtain an interpolation sample

[0076]

[0077] wherein ε ~ U[0, 1] is a random number satisfying uniform distribution between 0 and 1.

[0078] Note that the interpolation sample is a completely different concept from the interpolation sample obtained by the interpolation generation method of the present application.

[0079] The real sample and the sample pair (x, c) composed of the corresponding condition variable are input into the discriminator, and the Wasserstein loss D(x, c) is calculated according to the discrimination result of the real sample by the discriminator; the generated sample and the sample pair (x g , c) composed of the corresponding condition variable are input into the discriminator, and the Wasserstein loss D(x g , c) is calculated according to the discrimination result of the generated sample by the discriminator; the gradient penalty loss of the discriminator is calculated according to the interpolation sample wherein is the gradient of the discriminator D with respect to , and represents the Euclidean norm of the gradient, that is, the length of the gradient, and ideally the value is 1, so measures the deviation of the gradient, and the gradient penalty weight coefficient λ is introduced to control the importance of the gradient penalty term in the overall loss function.

[0080] The above three items constitute the discriminator loss of the conditional time series deep generative adversarial network module, and the discriminator loss is calculated as follows:

[0081]

[0082] For the generator, the training target is to generate a realistic real sample G(z, c) under the corresponding fault degree by using the noise sample pair (z, c) to "fool" the discriminator, that is, to train in the direction of the gradient of the output value of D(G(z, c), c) increasing.

[0083] The generator loss is calculated as follows:

[0084]

[0085] (3) Generation sample condition mutual information maximization module

[0086] ​The time-series deep generative adversarial network using the conditional information generates fault samples, and in an ideal case, can generate generated samples similar to real samples under a specific fault degree. However, whether the fault degree of the generated fault sample is consistent with the condition variable input as the conditional control information at the input end of the generator is not well constrained. In order to solve this problem, the application further introduces a generated sample conditional mutual information maximization module to construct a degradation trend reconstruction network, which takes the generated fault sample as input to reconstruct the degradation trend condition variable contained in the sample, and indirectly imposes a mutual information maximization constraint between the generated fault sample and the input condition variable. The principle of the module is as shown in Figure 6

[0087] For the conditional information fault sample generation task, we hope that the more information of the input condition variable contained in the generated fault sample, that is, the degradation information contained in the input condition variable is maximally retained in the generated sample. Mutual information can measure the dependence between two random variables:

[0088] I(X;Y) = H(X) - H(X|Y)

[0089] Therefore, the mutual information I(c; x g ) = I(c; G(z, c)) between the input condition variable and the generated fault sample is considered to measure the degree of retention of this degradation information, and is defined as conditional mutual information. Therefore, in addition to the loss function of the generator and the discriminator, the direction of maximizing this conditional mutual information needs to be additionally considered in the training process. However, it is very difficult to directly calculate I(c; G(z, c)), so an auxiliary distribution Q(c|x) is constructed to give a lower bound of the conditional mutual information:

[0090]

[0091] At this time, the goal of conditional mutual information maximization is to maximize L I (G, Q). At this time, the network Q is constructed to predict the mean and standard deviation of the condition variable. Assuming that the condition variable conforms to a normal distribution, for a random variable with a mean of μ and a standard deviation of σ, the probability density function is:

[0092]

[0093] To calculate the first part of L I (G, Q) is to calculate log p(c), that is Assuming that the mean and standard deviation of the condition variable output by the network Q are Q(x) μ and Q(x) σ , the conditional mutual information maximization optimization goal can be written as:​

[0094]

[0095] Since the distribution of the condition variable c is fixed, max L I (G,Q) is equivalent to:

[0096]

[0097] If the influence of Q(x) σ is not considered, further, the optimization objective max L I (G,Q) is equivalent to:

[0098] min(c-Q(x) μ ) 2

[0099] Therefore, the output of the network Q is a vector of the same dimension as the condition variable c. In the actual training process, in addition to the generator and discriminator loss of the conditional time-series deep generative adversarial network module, an additional generation sample conditional mutual information maximization loss is added:

[0100] L DE = MSE(q c ,c) = MSE(Q(x g ),c) = E[(Q(x g )-c) 2 ]

[0101] Where MSE is the mean square error function.

[0102] After introducing the generation sample conditional mutual information maximization loss based on the time-series deep generative adversarial network, the generated samples have the following three characteristics:

[0103] (11) The consistency between the generated fault samples and the real fault samples is high, which is guaranteed by the basic generative adversarial loss;

[0104] (12) The generated fault samples are matched with the input degradation trend condition variable under the control of the degradation trend condition variable of a specific fault degree, which is guaranteed by the conditional loss term in the generative adversarial loss, and also preliminarily restricts the fault degree of the generated samples to be consistent with the expectation;

[0105] (13) The generated samples retain the information in the input degradation trend condition variable to the greatest extent, which is guaranteed by the conditional mutual information maximization loss, and further improves the consistency between the fault degree of the generated samples and the expected fault degree.

[0106] For example, the XJTU-SY rolling bearing accelerated life test data set is used to verify the present application. Specifically as follows:

[0107] XJTU-SY bearing accelerated life test data set description:

[0108] The test rig used to collect the XJTU-SY bearing accelerated life test data set (hereinafter referred to as the data set) is composed of an alternating current motor, a motor speed controller, a rotating shaft, a support bearing, a hydraulic loading system, and a test bearing. The adjustable working conditions mainly include radial force and rotating speed, where the radial force is generated by the hydraulic loading system and acts on the bearing seat of the test bearing, and the rotating speed is set and adjusted by the rotating speed controller of the alternating current motor. The bearing model used in the test is LDKUER204, and three working conditions are set as shown in Table 1.

[0109] Table 1 XJTU-SY data set working condition information

[0110]

[0111] An acceleration sensor is fixed on the measured bearing seat, and the sampling frequency is 25.6 kHz. During the test, sampling is performed every 1 minute, and each sampling lasts for 1.28 seconds. To verify the effectiveness of the method proposed in this paper under different working conditions, data of one bearing is selected under each of the three working conditions for case analysis in this chapter. The three bearings are numbered as "Bearing1_3", "Bearing2_3", and "Bearing3_4". At the same time, the fault positions of the three bearings are the outer ring, the retainer, and the inner ring, respectively, which also verifies the effectiveness of the method under different fault modes.

[0112] Partial failure degree sample missing data set construction:

[0113] In order to simulate the sample missing under partial failure degree using the full life accelerated degradation test data, in this case, the method of setting anchor points in the full life data is used, taking the samples near the anchor points as the existing samples, and the other samples as the missing samples. The so-called anchor point is the known failure degree of the available failure sample in the full life data.

[0114] Taking bearing 1_3 as an example, it is assumed that only the failure samples at three failure degrees in the full life of the bearing are known, i.e. the number of anchor points is 3. Since the degradation process is stable, it is considered that the failure degree does not change in a short time, i.e. the failure degree does not change within a window of 4 sampling ranges before and after the anchor point. The data in this window can be used as the failure sample under the failure degree corresponding to this anchor point. For the samples other than the three anchor points, they are masked and treated as missing samples under unknown failure degrees, which are generated by using the method proposed in this paper. The schematic diagram of anchor point setting and masking treatment of the full life data of the bearing is as follows:Figure 7 as shown.

[0115] Test method and model parameter setting:

[0116] (101) Anchor point setting method

[0117] In this embodiment, anchor points are set in the original full-life accelerated test data in an equal interval manner. Assuming that a section of the bearing showing a clear degradation trend contains N original samples, and the number of anchor points to be set is k, and assuming that for a certain anchor point, s samples before and after the anchor point can be considered as not changing in failure degree, the calculation formula of the original sample interval contained by the jth anchor point is:

[0118]

[0119] wherein, represents the floor function.

[0120] Therefore, there are s original samples at each anchor point as known failure degree fault samples. In the full-life data, the positions other than the anchor point samples are unknown failure degree, and the corresponding fault samples are missing.

[0121] (102) Original data processing method

[0122] For each anchor point, each original sample contains 1.28*25600=32768 data points. Using a sliding window with a length of 1000 and a step size of 200, the original sample is cut to obtain the sample for model training. Therefore, after cutting, the length of each fault sample is 1000 data points. In this case, the number of anchor points k is set to 3, and the number of samples s considered not to change in failure degree before and after the anchor point is 4.

[0123] (103) Model parameter setting

[0124] In this case, RMS is selected as the degradation trend condition variable, which is used as the condition input of the generator to control the failure degree of the generated sample. RMS has the advantages of simple calculation, monotonicity and good trend, and is a health factor feature commonly used to represent the degradation state of equipment. In order to avoid the condition control effect of the model being masked by high-dimensional noise due to low dimension, the single-dimensional degradation trend condition variable is mapped to 5 dimensions, and then input to the generator together with the noise vector; correspondingly, before inputting the sample to the discriminator, the single-dimensional degradation trend condition variable also needs to be mapped to 5 dimensions. This condition variable is mapped from 1 dimension to 5 dimensions by setting a fully connected network layer in the model.

[0125] In this case, for the generator, discriminator and degradation trend reconstruction network in CRGAN-MDTR, the structural hyperparameters adopted are as follows:Figure 8 where Concatenate denotes the vector concatenation layer, BiLSTM represents the bidirectional LSTM layer, and Linear represents the full connection layer. All BiLSTM layers use the tanh activation function; all Linear layers use the LeakyReLU activation function except the output layer; the output layer Linear layer has no activation function and directly uses the calculation result after linear mapping.

[0126] The fault sample generation result is as follows:

[0127] According to the above parameters, the model is constructed, training is carried out, and the trained generator is used for fault sample generation. The generated samples are visually analyzed, and the results are as follows.

[0128] (1001) Anchor fault sample generation result

[0129] First, the samples at the anchor points in the training set are generated. The model can generate known samples in the training set, which is a prerequisite for generating samples under unknown fault conditions. Therefore, the generation effect of known samples is first analyzed. The RMS values of the existing samples at the three anchor points of each bearing are calculated as the degradation trend condition variables input into the generator; at the same time, random noise is sampled from the Gaussian distribution and input into the generator together with the degradation trend condition variables to generate fault samples at the anchor points. The time domain and frequency domain signals of the generated samples and the real samples are as shown in Figure 9 .

[0130] The results show that at the anchor points of the three bearings, the consistency of the generated fault samples and the real fault samples in the time domain and the frequency domain is very high, indicating that the CRGAN-MDTR model successfully learns the fault data distribution at each anchor point. When generating anchor samples, the noise vectors input into the generator are randomly sampled from the Gaussian distribution, and there is no significant difference between the noise vectors. The only difference is the condition variable; under this condition, the model can give different sample generation results for different condition variables, which is consistent with the real situation, further indicating that the model successfully learns the mutual dependence between the degradation trend condition variable and the fault sample.

[0131] (1002) Missing fault sample generation result

[0132] Further, the trained generator is used to generate missing fault samples under unknown fault levels. To show the real degradation trajectory of the bearing, the degradation trend condition variable for controlling the fault level of the generated samples is obtained by RMS calculation of the samples in the original accelerated life test data set. Since the length of each generated sample is 1000, the length of each sample in the original degradation data is also truncated to 1000 points for easy visualization comparison. The generated results of the fault samples in the full life degradation process of the three bearings are shown in Figure 10 .

[0133] The results show that the amplitudes of the original time domain signals of the generated samples and the real samples can basically match, and the RMS trends of the samples are basically consistent. In the generated results of Bearing2_3, the deviation between the RMS of some positions and the RMS of the real samples is relatively large, which may be because the degradation amplitude of this bearing is relatively small compared with the other two bearings, and the degradation trend is greatly affected by random fluctuations, which to some extent interferes with the learning of the CRGAN-MDTR on the distribution of the fault data in the degradation process.

[0134] The generated samples and the real samples in the full life degradation process are subjected to FFT transformation to obtain the frequency spectrum of the fault samples, and are visualized in the form of heat map, as shown in Figure 11 .

[0135] It can be seen that the similarity between the generated fault sample spectrum in the degradation process and the real fault sample spectrum is very high. Even under unknown fault levels without real samples for learning, the model can generate very real fault samples, which shows that the CRGAN-MDTR has the ability to interpolate and generate missing fault samples under unknown fault levels. Unlike fault sample generation under a single fault level, fault sample generation in a continuous degradation process not only requires the model to learn the static characteristics of the fault samples, but also to learn the dynamic characteristics exhibited by the samples as the fault level changes. From the results, it can be seen that as the fault level deepens, the model can restore the strength changes of the key frequency components; when the fault level changes to a new frequency component in the spectrum, the model can also restore the appearance of this new frequency component.

[0136] Fault diagnosis test:

[0137] The purpose of carrying out interpolation and generation of missing fault samples under unknown fault levels is to improve the performance of the fault diagnosis model under the condition of sample missing at partial fault levels. Therefore, directly comparing the performance improvement effect of the fault diagnosis model generated by different methods is the most direct way to verify the performance of the proposed fault sample generation method.

[0138] In this case, the training set TrainA with sample missing, the full life sample test set Test and the interpolation training set TrainB with generated samples under unknown fault degree are organized as follows:

[0139] 1) Training set TrainA with missing (normal samples + anchor point fault samples): Load the original samples of the three bearings at the three anchor point ranges respectively, cut them into short fault samples with a sliding window of length 1000 and step 200, and label the samples taken from the three bearings as 1, 2 and 3 respectively, representing three different fault modes of outer ring, cage and inner ring. Take one original sample from the fault-free stage of each of the three bearings, cut it into a short sample with a sliding window, and label it as 0, representing the normal state.

[0140] 2) Full life sample test set Test (normal samples + full degradation process fault samples): Load all original samples in the degradation trajectories of the three bearings showing obvious degradation trends, cut them into short fault samples with a sliding window, and label the samples taken from the three bearings as 1, 2 and 3 respectively; take the same number of original fault-free samples from the fault-free stage of each of the three bearings as the original samples contained in the degradation trajectory, cut them into short samples with a sliding window, and label them as 0.

[0141] 3) Interpolation training set TrainB (normal samples + anchor point fault samples + anchor point interpolation generated fault samples): First, load all the samples and labels of TrainA. Then, for each bearing, calculate the average RMS value of all anchor point fault samples, interpolate 100 points between the RMS values of the two anchor points uniformly as the degradation trend condition variable of the input generator for controlling the fault degree of the generated samples; for each interpolation point, randomly sample n noise vectors to generate n fault samples; label the generated samples of the three bearings as 1, 2 and 3 respectively. Finally, randomly select the same number of normal samples from the normal samples currently contained in TrainB as the number of generated fault samples, and copy them to ensure that the number of normal samples and fault samples under each fault mode in the dataset is balanced, eliminating the impact of dataset imbalance on the diagnosis results.

[0142] After the above datasets are constructed, the SVM model is trained with the datasets. Train with TrainA and test with Test to obtain the test set diagnosis accuracy, which can be considered as the benchmark condition for fault sample missing under partial fault degree during model training; Train with TrainB and test with Test, and set n in TrainB to different values to analyze and study the improvement effect of generated fault samples on diagnosis performance. The SVM test diagnosis accuracy results under the benchmark condition of fault sample missing (n = 0) and under different values of n are shown in Table 1. Figure 12

[0143] ​The results show that when a small amount of generated samples under unknown fault degrees are added to the training set with missing samples, the diagnostic accuracy of the SVM slightly decreases compared to when no generated samples are added, and the performance returns to the level before adding generated samples when the number of generated samples under each unknown fault degree reaches 500, because when the number of generated samples under each unknown fault degree is small, imbalance is formed between the samples under known fault degrees and the generated samples, which interferes with the learning of the model; when the number of generated samples continues to increase, the diagnostic performance is positively optimized, and the greater the number of generated samples, the higher the degree of improvement in the diagnostic model accuracy. This shows that in the fault diagnosis task under the condition of missing fault samples under unknown fault degrees, the CRGAN-MDTR model proposed in the present application can effectively generate missing fault samples, and by adding generated samples to the training set with missing samples, the performance of the fault diagnosis model can be improved in combination with existing samples.

[0144] Although the present application has been described in detail above, the present application is not limited thereto, and those skilled in the art can make various modifications according to the principles of the present application. Therefore, any modifications made in accordance with the principles of the present application should be understood to fall within the scope of the present application.

Claims

1. A method for interpolating missing samples under unknown fault severity, characterized in that: include: Extracting degradation trend condition variables capable of characterizing the degree of equipment failure from each fault sample under some known fault degrees, and constructing a training sample set using the degradation trend condition variables of each fault sample; Sampling the same number of noise vectors as the fault sample pairs in the training sample set from a predetermined prior noise distribution, and concatenating the noise vectors into a matrix by row; Copy and splice the degradation trend condition variables of the fault sample pair into the noise vector matrix so that the condition variables attached to the fault sample and the noise vector are the same, and use the noise vector matrix spliced ​​with the degradation trend condition variables of the fault sample pair as the training noise vector; Constructing a CRGAN-MDTR network model comprising a generator network, a discriminator network, and a degradation trend reconstruction network, and inputting the training noise vector into the generator network of the CRGAN-MDTR network model to obtain a generated sample, and inputting the generated sample into the degradation trend reconstruction network of the CRGAN-MDTR network model to obtain a reconstructed degradation trend conditional variable; Calculating interpolated samples based on the fault sample pairs in the training sample set and the generated samples, and calculating the discriminator loss using the fault sample pairs in the training sample set, the generated samples, and the interpolated samples; Calculating the generator loss based on the discriminator's discrimination result on the generated sample; According to the reconstruction results of the degradation trend conditional variables contained in the generated samples by the degradation trend reconstruction network, the conditional mutual information maximization loss of the generator network and the degradation trend reconstruction network is calculated; Based on the discriminator loss, the generator loss, and the conditional mutual information maximization loss, optimizing and updating the discriminator network, the generator network, and the degradation trend reconstruction network in sequence until the losses converge, completing the CRGAN-MDTR network model training and obtaining a trained CRGAN-MDTR network model. The trained generator network is then extracted from the trained CRGAN-MDTR network model. Obtaining a degradation trend condition variable under a target fault degree, and inputting the degradation trend condition variable under the target fault degree into the trained generator network to generate fault samples whose fault degree corresponds to the degradation trend condition variable under the target fault degree, thereby achieving interpolation generation of missing fault samples under an unknown fault degree; Among them, the CRGAN-MDTR refers to a conditional temporal generative adversarial network model that maximizes the degradation trend retention rate; the CRGAN-MDTR is based on the CGAN architecture. In terms of training structure, the temporal LSTM unit is introduced, and the LSTM unit with the ability to model and mine temporal features replaces the fully connected unit in the basic CGAN architecture. In terms of optimization objectives, the Wasserstein loss and gradient penalty are introduced, and the Wasserstein distance loss replaces the JS divergence loss in the basic CGAN.

2. The method according to claim 1, characterized in that The step of extracting degradation trend condition variables capable of characterizing the degree of equipment failure from each fault sample under some known fault degrees, and constructing a training sample set using the degradation trend condition variables of each fault sample includes: Extracting multiple features that can characterize the degree of equipment failure from each fault sample, and combining the multiple features that can characterize the degree of equipment failure to form a degradation trend condition variable corresponding to the fault sample; A fault sample pair with conditional information is formed by concatenating each fault sample with the degradation trend condition variable extracted from the fault sample, and all fault sample pairs are arranged in order from small to large fault severity; All fault sample pairs arranged in order of fault severity from small to large are used as training sample sets.

3. The method according to claim 2, characterized in that Generating a fault sample having a fault degree corresponding to the degradation trend condition variable under the target fault degree by inputting the degradation trend condition variable under the target fault degree into the trained generator network comprises: Sampling a noise vector from a Gaussian distribution, and obtaining a target noise vector by splicing a degradation trend condition variable under the target fault degree onto the noise vector; The target noise vector and the degradation trend condition variable under the target fault degree are inputted into the trained generator network to obtain a fault sample whose fault degree corresponds to the degradation trend condition variable under the target fault degree.

4. The method according to claim 3, characterized in that After obtaining the fault sample with the fault degree corresponding to the degradation trend condition variable under the target fault degree, the method further includes: By interpolation, the degradation trend condition variables under unknown fault degree are calculated between the degradation trend condition variables extracted from the existing fault samples, thereby completing the interpolation generation of missing fault samples under unknown fault degree.

5. A device for interpolating missing samples under unknown fault levels, characterized in that: include: A training sample set construction module is used to extract degradation trend condition variables that can characterize the degree of equipment failure from each fault sample under some existing known fault degrees, and to construct a training sample set using the degradation trend condition variables of each fault sample; and sampling the same number of noise vectors as the fault sample pairs in the training sample set from a predetermined priori noise distribution, and splicing the noise vectors into a matrix by row; copying and splicing the degradation trend condition variables of the fault sample pairs into the noise vector matrix so that the condition variables accompanying the fault samples and the noise vectors are the same, and using the noise vector matrix spliced ​​with the degradation trend condition variables of the fault sample pairs as the training noise vector; A model construction and training module is used to construct a CRGAN-MDTR network model including a generator network, a discriminator network, and a degradation trend reconstruction network, and obtain generated samples by inputting the training noise vector into the generator network of the CRGAN-MDTR network model, and then inputting the generated samples into the degradation trend reconstruction network of the CRGAN-MDTR network model to obtain reconstructed degradation trend conditional variables; According to the fault sample pairs in the training sample set and the generated samples, interpolated samples are calculated, and the discriminator loss is calculated using the fault sample pairs in the training sample set, the generated samples and the interpolated samples; the generator loss is calculated according to the discrimination result of the discriminator on the generated samples; according to the reconstruction result of the degradation trend conditional variables contained in the generated samples by the degradation trend reconstruction network, the conditional mutual information maximization loss of the generator network and the degradation trend reconstruction network is calculated; based on the discriminator loss, the generator loss and the conditional mutual information maximization loss, the discriminator network, the generator network and the degradation trend reconstruction network are optimized and updated in sequence until the loss converges, and the CRGAN-MDTR network model training is completed to obtain a trained CRGAN-MDTR network model, and the trained generator network is extracted from the trained CRGAN-MDTR network model; a sample interpolation generation module, configured to obtain a degradation trend condition variable under a target fault degree, and input the degradation trend condition variable under the target fault degree into the trained generator network to generate fault samples whose fault degree corresponds to the degradation trend condition variable under the target fault degree, thereby realizing interpolation generation of missing fault samples under an unknown fault degree; Among them, the CRGAN-MDTR refers to a conditional temporal generative adversarial network model that maximizes the degradation trend retention rate; the CRGAN-MDTR is based on the CGAN architecture. In terms of training structure, the temporal LSTM unit is introduced, and the LSTM unit with the ability to model and mine temporal features replaces the fully connected unit in the basic CGAN architecture. In terms of optimization objectives, the Wasserstein loss and gradient penalty are introduced, and the Wasserstein distance loss replaces the JS divergence loss in the basic CGAN.

6. The device according to claim 5, characterized in that The training sample set construction module is specifically used to extract multiple features that can characterize the degree of equipment failure from each fault sample, and combine the multiple features that can characterize the degree of equipment failure to form a degradation trend condition variable corresponding to the fault sample; A fault sample pair with conditional information is formed by splicing each fault sample with the degradation trend condition variable extracted from the fault sample, and all fault sample pairs are arranged in order from small to large fault severity; all fault sample pairs arranged in order from small to large fault severity are used as training sample sets.

7. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of a method for interpolating and generating missing samples under an unknown fault degree according to any one of claims 1 to 4 are implemented.

Images