A method for generating unknown working condition combined fault samples based on multi-attribute factor decoupling

Abstract

The application discloses a kind of unknown working condition combination fault sample recombination generation methods based on multi-attribute factor decoupling, comprising: on the basis of the fault sample under the condition of existing part known working condition combination, using the conditional distribution learning ability of deep generation model, in turn carry out original sample reconstruction, same attribute sample factor once interchanging reconstruction, different attribute sample factor twice interchanging reconstruction training, store the information related to each working condition attribute in the fault sample in the specific area of the non-overlapping hidden space vector, as the factor related to each working condition attribute, realize the decoupling of multi-working condition factor;Further, from the existing sample containing the shared attribute of target unknown working condition combination, carry out compression representation, factor extraction and cross recombination, complete the recombination generation of missing fault sample under unknown working condition, realize the diagnosis ability improvement of fault diagnosis model when facing complex variable working condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault data generation, in particular to a method for generating unknown working condition combination fault samples by decoupling multiple attribute factors. BACKGROUND

[0002] Working conditions can be reflected by multiple parameters, i.e., multiple working condition attributes. The combination of different working condition attribute levels determines the actual working condition of a product. The variable working condition problem is a popular research topic in the field of fault diagnosis. Existing research shows that the feature distribution of fault samples will change under different working conditions. Therefore, the difficulty of fault diagnosis under variable working condition conditions is greatly increased. The variable working condition characteristics of a product require that the fault diagnosis model has the anti-interference ability to working condition changes, i.e., the diagnosis model can accurately and quickly realize fault diagnosis regardless of the working condition in which the equipment is running. This requires that the fault diagnosis model training set contains fault samples under various working condition conditions. However, it is difficult to fully cover all combinations of working condition attributes under test or actual running conditions, resulting in the lack of fault samples under some unknown working conditions in the diagnosis model training set, which further affects the performance of the fault diagnosis model. SUMMARY

[0003] The present application provides a method for generating unknown working condition combination fault samples by decoupling multiple attribute factors, in order to solve the technical problem that the performance of the fault diagnosis model is affected due to the lack of some working condition combination fault samples in the multi-working condition attribute fault sample.

[0004] The present application provides a method for generating unknown working condition combination fault samples by decoupling multiple attribute factors, in order to solve the technical problem that the performance of the fault diagnosis model is affected due to the lack of some working condition combination fault samples in the multi-working condition attribute fault sample.

[0005] Preferably, the constructing the shared attribute sample joint graph using the respective working condition attributes of each failure sample in the failure sample set comprises: obtaining failure mode and working condition combination attributes of each failure sample in the failure sample set; dividing the failure sample set according to the failure mode to obtain a plurality of first failure sample sets; dividing the respective first failure sample sets according to the respective working condition combination attributes to obtain a plurality of second failure sample sets; and constructing the shared attribute sample joint graph using the plurality of second failure sample sets of each failure sample in the failure sample set.

[0006] Preferably, the constructing the shared attribute sample joint graph using the plurality of second failure sample sets of each failure sample in the failure sample set comprises: taking each failure sample in the failure sample set as a node, taking a shared working condition attribute relationship between two failure samples as an edge, and taking an attribute label corresponding to the shared working condition attribute as a value of the edge to construct the shared attribute sample joint graph.

[0007] Preferably, the training the encoding-decoding structure deep generative model using the shared working condition attribute relationships between the respective failure samples in the shared attribute sample joint graph comprises: sequentially performing original sample compression reconstruction, same-attribute sample factor one-time exchange reconstruction, and different-attribute sample factor two-time exchange reconstruction training on the encoding-decoding structure deep generative model using the shared working condition attribute relationships between the respective failure samples in the shared attribute sample joint graph to obtain the encoding-decoding structure deep generative model having the function of decoupling the hidden variable factors corresponding to the plurality of working condition attributes.

[0008] Preferably, the original sample compression reconstruction training of the encoding-decoding structure deep generative model comprises: an original sample compression reconstruction loss using the shared working condition attribute relationships between the respective failure samples in the shared attribute sample joint graph and taking a mean square error between each original failure sample and a reconstruction sample obtained by encoding and decoding of the original failure sample as the original sample compression reconstruction loss L rec .

[0009] Preferably, the same-attribute sample factor one-time exchange reconstruction training of the encoding-decoding structure deep generative model comprises:

[0010] Original sample encoding process: z = E(x), z o = E(x o )

[0011] Hidden variable factor exchange process:

[0012] Exchange hidden variable decoding process:

[0013] Calculate the reconstruction loss of samples with the same attribute after a single factor swap:

[0014] Where, z, z o It is composed of sample x, x o Latent variables derived from encoding; z s , It is a latent variable z, z o The new latent variable obtained after swapping the j-th factor; It is caused by the latent variable z s , The generated sample obtained from decoding.

[0015] Preferably, the deep generative model of the encoder-decoder structure undergoes heterogeneous sample factorization reconstruction training, which includes:

[0016] Select fault sample pairs Each fault sample in the fault sample pair is then input into the encoder for encoding to obtain the latent variables:

[0017] z = E(x),

[0018] For the latent variable z and By swapping the factors corresponding to attribute j, we obtain the swapped latent variables:

[0019]

[0020] The latent variable z will be swapped s , The input is decoded to obtain the generated sample after one swap:

[0021]

[0022] The generated samples after another swap and The data is fed into the encoder to obtain the latent variables of the secondary encoding.

[0023]

[0024] Latent variables for secondary encoding and The factors corresponding to attribute j are interchanged again to obtain the latent variables of the second interchange:

[0025]

[0026] Double-cross the hidden variables Input to the decoder, decode to obtain the generated sample after the second transposition:

[0027]

[0028] The deviation between the generated sample pair after the second exchange of the two attributes and the original sample pair is measured using the mean square error as the second exchange of the attribute-specific sample factor reconstruction loss:

[0029]

[0030] Preferably, the recombination generation of the missing fault sample under the unknown working condition combination by inputting the fault sample set combined with the known working condition of the same working condition attribute into the deep generative model with the encoding-decoding structure having the function of decoupling the corresponding latent variable factor of multiple working condition attributes includes: using the encoder network in the deep generative model with the encoding-decoding structure having the function of decoupling the corresponding latent variable factor of multiple working condition attributes to encode the fault sample combined with the known working condition into the latent variable related to the working condition factor, and obtain the latent variable containing unknown working condition combination information through cross exchange between the latent variables; using the decoder network in the deep generative model with the encoding-decoding structure having the function of decoupling the corresponding latent variable factor of multiple working condition attributes to decode the latent variable containing unknown working condition combination information into a fault sample, thereby realizing the recombination generation of the missing fault sample under the unknown working condition combination.

[0031] Preferably, the use of the encoder network in the deep generative model with the encoding-decoding structure having the function of decoupling the corresponding latent variable factor of multiple working condition attributes to encode the fault sample combined with the known working condition into the latent variable related to the working condition factor, and obtain the latent variable containing unknown working condition combination information through cross exchange between the latent variables includes: feeding all samples in the fault sample set combined with the known working condition of the same working condition attribute into the encoder network to extract the latent variable:

[0032] z (i) =E(x (i) ), for i = 1, …, m

[0033] Wherein, each latent variable is composed of corresponding m decoupling factors controlling m attributes: The factor controlling attribute i in the latent variable z (i) of each sample is extracted and recombined into a new latent variable as the latent variable containing unknown working condition combination information:

[0034]

[0035] Preferably, the decoder network in the deep generative model with the encoding-decoding structure having the function of decoupling the plurality of working condition attributes corresponding to the latent variable factor decodes the latent variable containing the unknown working condition combination information into a fault sample, so as to realize the missing fault sample reorganization generation under the unknown working condition combination.

[0036]

[0037] The advantages and positive effects of the present application are:

[0038] (1) A shared attribute sample connection graph construction method for multi-working condition attribute factor decoupling representation learning is proposed, the shared attribute information between the existing fault samples is structured expressed in the form of multi-graph, the data and attribute information source for the deep generative model training of the encoding-decoding structure are provided, and thus the decoupling of the plurality of working condition factors is realized only by using the reconstruction error loss.

[0039] (2) A multi-attribute factor decoupling and unknown working condition attribute combination sample reorganization generation method based on factor exchange reconstruction invariance is proposed by referring to the idea of gene controlling traits, the method of compressing and representing the information related to each working condition attribute into independent factors that do not overlap and do not affect each other in the latent variable is learned by the model in a supervised manner through three reconstruction generation losses, and the effectiveness and generation quality of the model when generating the samples not seen in the training set are improved by combining with the supplementary generation adversarial loss, so as to realize the missing fault sample reorganization generation under the unknown working condition combination.

[0040] (3) The principle and process effectiveness of the present application for realizing the multi-attribute factor decoupling and unknown attribute combination sample reorganization generation are verified and analyzed by combining with the artificial constructed simulation data set, through sample generation process visualization, factor decoupling representation visualization and other ways, and support is provided for the application of the present application in actual fault data.

[0041] (4) Combined with the multi-failure mode and multi-working condition attribute data set of the aviation high-speed bearing, the generated method of the missing fault sample under the unknown working condition combination is verified and analyzed through case, the working condition factor decoupling ability and the missing fault sample generation ability under the unknown working condition combination are verified through sample visualization, feature manifold space visualization and quantitative index calculation, and the sample generation result of the method is better than that of the similar method in qualitative and quantitative aspects, the superiority of the method in decoupling and generation tasks is verified, and the missing fault sample under the unknown working condition combination is recombined and generated by using the method, so that the performance of the diagnosis model can be significantly improved under the sample missing condition, and the decoupling representation and sample generation mechanism of the method are preliminarily explained through the method of convolution kernel visualization. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a schematic diagram of the unknown working condition combination fault sample recombination generation method based on multi-attribute factor decoupling provided by the application;

[0043] Figure 2 is a schematic diagram of the shared attribute sample connection diagram provided by the application;

[0044] Figure 3 is a schematic diagram of the simulation data set sample matrix provided by the application;

[0045] Figure 4 is a schematic diagram of the simulation data set original sample provided by the application;

[0046] Figure 5 is a schematic diagram of the simulation data set original sample spectrum provided by the application;

[0047] Figure 6 is a schematic diagram of the loss curve of the SIMDRN training process provided by the application;

[0048] Figure 7 is a schematic diagram of the original sample reconstruction generation result provided by the application;

[0049] Figure 8 is a schematic diagram of the attribute f1 same sample factor one-time exchange reconstruction generation result provided by the application;

[0050] Figure 9 is a schematic diagram of the attribute f2 same sample factor one-time exchange reconstruction generation result provided by the application;

[0051] Figure 10 is a schematic diagram of the no same attribute sample factor two-time exchange intermediate sample and reconstruction generation result provided by the application;

[0052] Figure 11It is the attribute f2 same sample factor dimension reduction visualization schematic provided by the application.

[0053] Figure 12 It is the attribute f1 same sample factor dimension reduction visualization schematic provided by the application.

[0054] Figure 13 It is the DIRG data set test bench schematic provided by the application.

[0055] Figure 14 It is the DIRG data set sample matrix schematic provided by the application.

[0056] Figure 15 It is the missing fault sample recombination generation result schematic under the condition of (100Hz, 1000N) provided by the application.

[0057] Figure 16 It is the missing fault sample recombination generation result schematic under the condition of (100Hz, 1800N) provided by the application.

[0058] Figure 17 It is the missing fault sample recombination generation result schematic under the condition of (200Hz, 1400N) provided by the application.

[0059] Figure 18 It is the missing fault sample recombination generation result schematic under the condition of (300Hz, 1000N) provided by the application.

[0060] Figure 19 It is the missing fault sample recombination generation result schematic under the condition of (300Hz, 1800N) provided by the application.

[0061] Figure 20 It is the DIRG data set generation sample and real sample dimension reduction visualization schematic provided by the application.

[0062] Figure 21 It is the DIRG data set same attribute sample clustering area visualization schematic provided by the application.

[0063] Figure 22 It is the flow chart of the unknown working condition combined fault sample recombination generation method based on multi-attribute factor decoupling provided by the application. DETAILED DESCRIPTION

[0064] It should be understood that the specific embodiments described herein are merely intended to explain the application and are not intended to limit the application. In the subsequent description, the suffixes such as "module", "component", or "unit" used to represent elements are only for the convenience of the description of the application, and have no specific meaning. Therefore, "module", "component", or "unit" can be mixedly used.

[0065] The application draws on the idea of gene controlling traits in genetics, and on the basis of existing partial known working condition combinations and fault samples (obtained by fault injection and the like), uses the conditional distribution learning ability of a deep generative model to successively carry out original sample reconstruction, same-attribute sample factor one-time exchange reconstruction and different-attribute sample factor two-time exchange reconstruction training, stores information related to each working condition attribute in the fault samples into specific areas in the hidden space vector that do not overlap with each other as factors related to each working condition attribute, realizes decoupling of multiple working condition factors, and further carries out compression representation, factor extraction and cross-recombination from existing samples containing shared attributes of a target unknown working condition combination to complete missing fault sample reconstruction generation under unknown working conditions, and realize the improvement of the diagnosis capability of a fault diagnosis model under complex variable working condition conditions.

[0066] According to the above description, the function realized by the fault sample reconstruction generation method provided by the application is as follows:

[0067] Known X ss ~p(x|u s ,v s ), X tt ~p(x|u t ,v t ), find: p(x|u s ,v t ), p(x|u t ,v s ). Wherein u i , v j represent the value level i of the working condition factor u and the value level j of the working condition factor v respectively; X ss represents the fault sample under the value level s of the working condition factors u and v, and X tt represents the fault sample under the value level t of the working condition factors u and v.

[0068] Since only a small amount of fault samples under known working condition combinations can be used for model training when fault samples are missing under unknown working condition combinations, the learning goal is to construct the independent conditional dependency relationship between the fault samples and multiple working condition attributes, and then generate the fault samples of the target working condition combination, and the fault sample reconstruction generation method provided by the application can express the independent conditional dependency relationship between the fault samples and the working condition attributes through factor decoupling, so the method provided by the application has good applicability to the missing condition of fault samples under unknown working condition combinations.

[0069] Figure 22 is a flowchart of a fault sample reconstruction generation method for unknown working condition combinations based on multiple attribute factor decoupling provided by the application, as Figure 22As shown, comprising: step S101: obtaining a fault sample set of known working condition combinations, and constructing a shared attribute sample joint graph using the working condition combination attributes of each fault sample in the fault sample set; step S102: constructing a deep generative model of an encoding-decoding structure, and training the deep generative model of the encoding-decoding structure using the shared working condition attribute relationships between the fault samples in the shared attribute sample joint graph, to obtain a deep generative model of an encoding-decoding structure having a function of decoupling hidden variable factors corresponding to multiple working condition attributes; step S103: obtaining an unknown working condition combination to be generated, and selecting a fault sample set of a known working condition combination same as the working condition attributes of the unknown working condition combination to be generated; step S104: obtaining a reorganized generated sample of a missing fault sample under the unknown working condition combination by inputting the fault sample set of the known working condition combination same as the working condition attributes into the deep generative model of the encoding-decoding structure having the function of decoupling hidden variable factors corresponding to multiple working condition attributes.

[0070] Further, the construction of the shared attribute sample joint graph using the respective working condition attributes of each fault sample in the fault sample set comprises: obtaining fault mode and working condition combination attributes of each fault sample in the fault sample set; dividing the fault sample set according to the fault modes to obtain a plurality of first fault sample sets; dividing the respective first fault sample sets according to the respective working condition combination attributes to obtain a plurality of second fault sample sets; and constructing a shared attribute sample joint graph using the plurality of second fault sample sets of each fault sample in the fault sample set.

[0071] More specifically, the construction of the shared attribute sample joint graph using the plurality of second fault sample sets of each fault sample in the fault sample set comprises: taking each fault sample in the fault sample set as a node, taking a shared working condition attribute relationship between two fault samples as an edge, and taking an attribute label corresponding to the shared working condition attribute as a value of the edge, to construct a shared attribute sample joint graph.

[0072] Further, the training of the deep generative model of the encoding-decoding structure using the shared working condition attribute relationships between the fault samples in the shared attribute sample joint graph to obtain a deep generative model of an encoding-decoding structure having a function of decoupling hidden variable factors corresponding to multiple working condition attributes comprises: sequentially performing original sample compression reconstruction, same-attribute sample factor one-time exchange reconstruction, and different-attribute sample factor two-time exchange reconstruction training on the deep generative model of the encoding-decoding structure using the shared working condition attribute relationships between the fault samples in the shared attribute sample joint graph, to obtain a deep generative model of an encoding-decoding structure having a function of decoupling hidden variable factors corresponding to multiple working condition attributes.

[0073] The original sample compression reconstruction training of the deep generative model of the encoding-decoding structure includes: the original sample compression reconstruction loss utilizes the shared working condition attribute relationship between each fault sample in the shared attribute sample joint graph, and calculates the mean square error between each original fault sample and the reconstruction sample obtained by encoding-decoding as the original sample compression reconstruction loss L rec ,

[0074]

[0075] In the formula, L rec is the original sample compression reconstruction loss, is an expectation function, is a compression reconstruction sample, and x is an original sample.

[0076] The same-attribute sample factor one-time exchange reconstruction training of the deep generative model of the encoding-decoding structure includes:

[0077] The original sample encoding process: z=E(x), z o =E(x o )

[0078] The hidden variable factor exchange process:

[0079] The exchanged hidden variable decoding process:

[0080] The same-attribute sample factor one-time exchange reconstruction loss is calculated:

[0081] The different-attribute sample factor two-time exchange reconstruction training of the deep generative model of the encoding-decoding structure includes:

[0082] Selecting a fault sample pair and inputting each fault sample in the fault sample pair into the encoder to obtain hidden variables:

[0083] z=E(x),

[0084] Exchanging the factors corresponding to the attribute j of the hidden variables z and to obtain exchanged hidden variables:

[0085]

[0086] Inputting the exchanged hidden variables z s , into the decoder to obtain the generated sample after one-time exchange:

[0087]

[0088] the first interchanged generated samples and are input into the encoder to obtain the second encoded latent variables:

[0089]

[0090] the second encoded latent variables and the factors corresponding to the attribute j are interchanged again to obtain the second interchanged latent variables:

[0091]

[0092] the second interchanged latent variables are input into the decoder to decode to obtain the second interchanged generated samples:

[0093]

[0094] the mean square error is used to measure the deviation between the pair of second interchanged generated samples and the original sample pair as the attribute different sample factor second interchanged reconstruction loss:

[0095]

[0096] Further, the method comprises: inputting the fault sample set combined with the known working condition attribute into the deep generative model having the function of decoupling the latent variable factors corresponding to multiple working condition attributes to obtain the recombined generated sample of the missing fault sample under the unknown working condition combination.

[0097] Specifically, the method comprises: inputting all samples in the fault sample set combined with the known working condition attribute into the encoder network to extract the latent variables:

[0098] z (i) = E(x (i) ), for i = 1, …, m

[0099] where each latent variable is composed of the corresponding m decoupled factors that control m attributes: extract the factor that controls attribute i in z (i) from each sample and reorganize it into a new latent variable as the latent variable containing unknown working condition combination information:

[0100]

[0101] Specifically, the decoder network in the deep generative model using the encoding-decoding structure having the latent variable factor decoupling function corresponding to multiple working condition attributes decodes the latent variable containing unknown working condition combination information into a fault sample, thereby realizing the reorganization and generation of missing fault samples under unknown working condition combinations, including: decoding in the input decoder network containing unknown working condition combination information to obtain a reorganization and generation sample of missing fault samples under unknown working condition combinations:

[0102]

[0103] The application provides an unknown working condition combination fault sample reorganization and generation method based on multi-attribute factor decoupling.

[0104] Step 1, divide existing fault samples into different sets according to fault modes and working conditions, for example, first divide bearing fault data into sets S1 = {outer ring fault}, S2 = {inner ring fault}, S3 = {rolling element fault} and the like according to fault modes, and then divide each set into sub-sets S 11 = {outer ring fault, rotating speed 100 Hz, load 1000 N}, S 12 = {outer ring fault, rotating speed 200 Hz, load 1200 N}, S 21 = {inner ring fault, rotating speed 100 Hz, load 1000 N}, S 22 = {inner ring fault, rotating speed 200 Hz, load 1200 N} and the like according to working condition combinations.

[0105] Step 2, construct a shared attribute sample connection graph according to working condition attributes of existing fault samples, and the shared attribute sample connection graph can structurally show the shared attribute relationship between existing samples, and the shared attribute relationship between existing samples is the basis for applying the application.

[0106] Step 3, a deep generative model of the encoding-decoding architecture is constructed and trained based on the swap invariance loss to realize the decoupling of the hidden variable factors corresponding to multiple working condition attributes. The swap invariance loss includes four parts, which are original sample compression reconstruction loss, same attribute factor swap reconstruction loss, different attribute factor swap reconstruction loss and supplementary generative adversarial loss.

[0107] Step 4, the existing fault samples are encoded into hidden variables related to working condition factors by using the encoder network in the trained generative model, the hidden variables containing unknown working condition combination information are obtained through cross swap between the hidden variables, and finally the hidden variables are decoded into fault samples by using the decoder network, so that the missing fault samples under unknown working condition combination are reorganized and generated.

[0108] Step 5, the fault diagnosis model is trained by using the fault samples under known working condition combination and the fault samples under unknown working condition combination generated by the method of the application, the quality of the generated samples is evaluated by fault diagnosis accuracy, and enhanced fault diagnosis under partial working condition sample missing is realized.

[0109] To solve the problem of missing fault samples under unknown working condition combination, the application proposes a swap invariance-based multi-attribute disentangled representation network (SIMDRN), which is a deep generative model of the encoding-decoding architecture. First, a shared attribute sample connection graph is constructed according to the attributes of the existing fault samples, then a deep generative model of the encoding-decoding architecture is constructed and trained based on the swap invariance loss to realize the decoupling of the hidden variable factors corresponding to multiple working condition attributes, and finally the trained encoder network is used to extract and reorganize the hidden variables containing unknown working condition factor combination from the existing fault samples, and the trained decoder network is used to realize the reorganization and generation of missing fault samples under unknown working condition combination. The method flow of the unknown working condition combination sample reorganization and generation based on multi-attribute factor decoupling proposed in the application is shown in Figure 1 .

[0110] Before describing the method, two important concepts involved in the application need to be defined first.

[0111] (1) Attribute: refers to a certain property possessed by a sample. For example, for a human face image class sample, the hairstyle, hair color, skin color, expression, and face orientation of the face contained in the image are all attributes of the sample. For example, for a rotating machinery fault sample, the fault mode, fault degree, speed, and load of the object during sample collection are also attributes of the sample. Image class samples exist in the form of RGB, and the attributes of the samples can be directly observed by the human eye. However, the above-mentioned sample attributes cannot be directly obtained by observation. However, whether it can be directly observed or not, different attributes of the sample are in different ways to act on the sample itself, so that samples with different attributes exhibit different properties.

[0112] (2) Factor: refers to a specific part of the hidden variable extracted from the original sample using a deep network. The factor is an abstract concept. The factor corresponding to a certain attribute of the sample is the part related to the attribute in the hidden variable containing all the necessary information of the original sample after the original sample is calculated by the deep network forward propagation.

[0113] According to the above definition, the attribute can be understood as the abstract way of human beings to the sample in the objective world, and the factor is the abstract way of the deep network to the sample in the high-dimensional data space. The sample is the medium between the attribute and the factor. For example, for the feature extraction process, different attribute values determine the morphology of the sample, and different sample morphologies determine the factors extracted by the deep network. Conversely, for the sample generation process, different factor values determine the morphology of the sample, and different sample morphologies represent different attributes of the sample.

[0114] The overall idea of the embodiment of the present application is to regard different working condition parameters (such as speed, load, etc.) as sample attributes, use a deep generative model with an encoder-decoder structure to learn a factor expression method that is mutually exclusive and mutually independent between attributes, and then generate missing fault samples under working condition attribute combinations not covered in the training set by constructing hidden variables of different factor combinations.

[0115] The precondition for applying the fault sample recombination generation method proposed in the present application is that the related fault mode and fault mechanism do not change in the process of changing various working condition attributes, that is, the influence of the change of working condition attributes on the fault sample remains within the range of quantitative change.

[0116] The method described in the present application comprises the following three implementation steps:

[0117] Step 1: Shared attribute sample connection graph construction

[0118] The shared attribute sample join graph refers to a graph model constructed by taking existing fault samples as nodes and taking the shared attribute relationship between the samples as edges. If two samples have the same attribute value (i.e., share the attribute), there is an edge between the nodes corresponding to the two samples in the graph, and the value of the edge is the label corresponding to the shared attribute. The shared attribute sample join graph describes the case where samples contain the same attribute, and is the input for subsequent training of a deep generative model with decoupling capability.

[0119] For an existing data set composed of n samples Each sample contains m attributes, and the attribute set possessed by all samples is Wherein represents the value of the jth attribute of the ith sample. The value of each attribute comes from a finite set, a j ∈A j For example, for a bearing fault sample expressed by a vibration signal, the attributes can include the fault mode, fault degree, rotating speed, load, etc. of the fault bearing during sample collection; and each attribute comes from a different value set, for example, A1={inner ring fault, outer ring fault,...}, A2={0.007 mm, 0.014 mm,...}, A3={100 Hz, 200 Hz,...}, and A4={1000 N, 2000 N,...}.

[0120] Using the above n samples and their attribute sets, a multi-graph M is constructed, which contains a node set [1,...,n]. For two different nodes i and k, i,k∈[1...n] and i≠k, there are a plurality of labeled edges connected between them, and the set of edges connecting i and k is:

[0121]

[0122] represents the label set corresponding to the shared attributes of samples i and k. A schematic diagram of a constructed shared attribute sample join graph is shown in Figure 2 .

[0123] In the multi-graph M, |M(i,k)| represents the number of edges connecting nodes i and k, i.e., the number of shared attributes between samples i and k. Based on the constructed multi-graph M, a covering relationship between a node set S and node i is defined as follows:

[0124] For a given node set and node If for each attribute value of i, there is at least one element in S that shares the attribute with i, S is said to cover i:

[0125]

[0126] The covering relation is not equivalent to the inclusion relation. When COVER(S, i) holds, it is possible that i∈S or These are two mutually exclusive cases. For the first case, i∈S, COVER(S, i) holds when S is some small set (e.g. COVER(S2, x2) in the figure above); even considering the most extreme case, S = {i}, COVER(S, i) = COVER({i}, i) is always true. However, in this case, the method proposed by the present application will degenerate to the most basic autoencoder model. Therefore, the present application is mainly concerned with the second case, i∉S, and COVER(S, i) holds, such as COVER(S1, x2) in the figure above. COVER(S, i) holds, such as COVER(S1, x2) in the figure above.

[0127] Step two, working condition factor decoupling based on interchange invariance

[0128] If the deep generative model can learn the decoupled representation of the original data, one factor in the latent variable controls the corresponding attribute of the generated sample, which is similar to the characteristics of genes controlling traits in genetics. If two chromosomes are exchanged when crossing over, the genotypes of the two exchanged chromosomes at a certain gene locus are the same, and the offspring produced by the two chromosomes will not mutate in the trait controlled by this gene. Inspired by this, the present application proposes a working condition factor decoupling method based on interchange invariance, which introduces a four-part sample reconstruction error loss to encourage the deep generative model of the encoding-decoding architecture to learn the decoupled representation factors stored in the latent variables during the compression and reconstruction of the original data, and thus enable the model to have the ability to recombine samples that meet specific attribute combinations. The four-part loss is described below.

[0129] (1) Original sample compression and reconstruction loss

[0130] This part of the loss is to hope that the model learns the compression representation of the original sample, so that the model can compress all the information in the original sample into the latent variables in the latent space, and then restore the latent variables to the original data space to complete the sample reconstruction. This process is achieved through an autoencoder, and in addition, the shared attribute sample connection graph M constructed in the previous section needs to be considered in the autoencoder.

[0131] The encoder and decoder network structure connected in series in the autoencoder is denoted as where the encoder network is The decoder network is For SIMDRN, the joint graph M needs to be further considered, that is, the encoder and decoder need to consider not only the samples and the latent variables themselves, but also the attribute sharing relationship between the samples. Therefore, the encoder and decoder networks in SIMDRN are respectively represented as:

[0132] Encoder E:

[0133] Decoder D:

[0134] That is, for the constructed shared attribute sample joint graph, the pair relationship between the sample pairs in the data set is represented. In the training process of SIMDRN, a batch of training samples X is sampled from the original sample set , and the shared attribute joint graph M of the batch of training samples is used to train the model, where X contains a batch of training samples, and M is a subgraph of the shared attribute joint graph of the batch of training samples. The detailed method of this step is as follows: a batch of training samples X is selected from all the samples , so that the shared attribute joint graph M of the training samples is a subgraph of the shared attribute joint graph of all the samples. The original sample compression reconstruction loss only utilizes the relationship of each sample in with itself, and the mean square error between each original sample and the reconstructed sample obtained by encoding-decoding is calculated as the loss, denoted as L rec , and the specific calculation process is as follows:

[0135]

[0136] In the formula, L rec is the original sample compression reconstruction loss, is the expectation function, is the compression reconstruction sample, and x is the original sample.

[0137] (2) Same attribute factor exchange reconstruction loss

[0138] This part of the loss is to hope that the model learns that if two samples share an attribute, then before decoding the respective latent variables, the factors corresponding to the shared attribute are exchanged, and the reconstructed sample obtained by decoding should be consistent with the original sample. This makes SIMDRN preliminarily decouple the factors corresponding to different attributes in the latent variable representation.

[0139] At this time, the structure of the autoencoder is changed, and a transformation is inserted before the output latent variable of E is sent to D for decoding, denoted as A, and the model structure at this time becomes In order to simplify the expression, a factor exchange operation is introduced, and A is merged into E.

[0140] Before training the SIMDRN, the latent variable Z = E(X, M) needs to be partitioned first. Assume that the latent variable z of dimension d is composed of m sub-vectors:

[0141] z = [g1, g2, …, g m ]

[0142] where i.e. the sum of the dimensions of all sub-vectors equals the dimension of the latent variable z. is a hyper-parameter of the model, Different settings of are different ways of partitioning the latent variable. i.e. the factors in the latent variable corresponding to different attributes.

[0143] Next, define the factor exchange operation:

[0144]

[0145] The above factor exchange operation takes two latent variables and an attribute label as input and returns two output latent variables with the same dimension as the input latent variables; in these two output latent variables, the factors corresponding to the input attribute label are exchanged with each other. For example:

[0146]

[0147] The positions marked by arrows are the positions of the exchanged factors. It can be seen that the second factors of z (1) and z (2) have completed the exchange operation.

[0148] Consider a set of samples S and a sample x, such that COVER(S, x) holds, and for all x o ∈ S, x ≠ x o According to the shared attribute join graph M, the sample pair (x, x o ) shares attribute j. Input both samples into E to encode and get the original latent variable; then exchange the factors in the original latent variable corresponding to attribute j to get the exchanged latent variable; input the exchanged latent variable into D to decode. If the latent variable factors are completely decoupled, then after exchange, the other attributes in the decoded generated sample should not be affected. Since the sample pair (x, x o ) shares attribute j, according to the attribute factor exchange invariance principle, the sample generated by decoding using the exchanged latent variable should be consistent with the original sample respectively. The mathematical expression of this process is as follows:

[0149] Original sample encoding process: z = E(x), z o = E(x o )

[0150] where z, z o is the latent variable encoded from sample x, x o .

[0151] Latent variable factor swapping process:

[0152] where z s , is the latent variable z, z o after the jth factor is swapped.

[0153] Swapped latent variable decoding process:

[0154] where is the generated sample decoded from latent variable z s , .

[0155] Same attribute sample swapping reconstruction loss

[0156] (3) Different attribute factor swapping reconstruction loss

[0157] This part of the loss is expected to learn that if the sample after factor swapping decoding has no real sample in the training set as supervision information, then the generated sample after decoding can be encoded-swapped-decoded again, that is, the generated sample can be restored to the original input sample, so as to obtain supervision information. This makes the decoupling ability of SIMDRN further enhanced, and has the processing ability of unknown samples.

[0158] If a pair of samples does not share an attribute, then the generated sample obtained by encoding, swapping the factor corresponding to this attribute, and decoding will have no corresponding real sample as a label for supervised learning. Even, in addition to this pair of samples, there may be no corresponding real sample in the entire training data set that can be used as supervision information. At this time, we draw lessons from the idea of cycle generation adversarial loss in CycleGAN, and do not use the method of directly supervised learning after encoding-factor swapping-decoding, but repeat the swapping process twice to indirectly obtain supervision information.

[0159] Given a pair of samples Randomly select an attribute j ~ U [1, …, m] from m attributes. Encode the pair of samples into the encoder respectively to obtain the latent variable z,

[0160] z = E(x),

[0161] Swap the jth factor of z and z Interchange the factors corresponding to attribute j, get the interchanged latent variable z s ,

[0162]

[0163] Send z s , into the decoder, get the generated sample after the first interchange:

[0164]

[0165] Send z and z into the encoder again, get the secondly encoded latent variable z

[0166]

[0167] Interchange z and z again, get the secondly interchanged latent variable z

[0168]

[0169] Finally, send z into the decoder, get the generated sample after the second interchange

[0170]

[0171] In this process, although z and z Since z does not share attribute j, there is no corresponding real sample in the training set that can be used as a label for supervised training; but after the second interchange, if the model can well decouple the factors corresponding to different attributes, then z should be restored to z At the same time, it also indirectly constrains the authenticity of the intermediate sample generated after the first interchange of the cross-attribute factors. Because if the intermediate sample is not a sample that meets the expected attribute combination, or the quality of the generated sample is very poor, then there will be a large difference between the sample generated after the second interchange and the original sample input. Therefore, the mean square error is used to measure the deviation between the pair of generated samples after the second interchange and the pair of original samples, as the cross-attribute sample interchange reconstruction loss:

[0172]

[0173] To realize the decoupling of working condition factors, the training optimization objective of SIMDRN is the weighted sum of the original sample compression reconstruction loss, the same attribute sample exchange reconstruction loss, and the different attribute sample exchange reconstruction loss:

[0174]

[0175] where θ represents the parameter set of the encoder and decoder networks in SIMDRN, λ1, λ2, and λ3 are the weight parameters of the three parts of the loss, arg min represents the parameter combination that can minimize θ, L rec is the original sample compression reconstruction loss, L swap is the same attribute factor exchange reconstruction loss, and L cycle is the different attribute factor exchange reconstruction loss. disengangle is the weighted sum of the three parts of the reconstruction loss.

[0176] (4) Supplementary generative adversarial loss

[0177] Although in L cycle , the constraint on the authenticity of the intermediate generated sample is realized, but this constraint is indirect, and it is more focused on whether the attributes of the intermediate sample are correct, and the constraint on the sample quality is not strong. Therefore, in order to further improve the quality of the intermediate sample in the secondary exchange process of different attribute factors, a supplementary generative adversarial loss is introduced here, and a discriminator is set to evaluate whether the intermediate sample is similar to the real sample. Here, the real sample used to optimize the supplementary generative adversarial loss does not necessarily come from the real sample whose attributes are consistent with the intermediate generated sample, but can come from the entire training set composed of all known fault samples. In other words, the supplementary generative adversarial loss focuses on making up for the lack of constraint on the quality of the intermediate generated sample in L cycle , rather than the constraint on the sample attributes.

[0178] The introduced discriminator is denoted as C, and E, the swap operation, and D in SIMDRN together constitute the generator of this process, denoted as G. The Wasserstein loss is used, and a gradient penalty term is added. The loss function is calculated as follows.

[0179] The sample generated by decoding after the first exchange of different attribute factors is and In the supplementary generative adversarial loss part, the attribute information of the two is not concerned, so and are uniformly denoted as The real sample x' is sampled from the existing fault sample set. The linear interpolation between and x' is calculated as follows: Then the loss of C is:

[0180]

[0181] where λ gp is the gradient penalty term, is the gradient operator.

[0182] The loss of G composed of E and D is:

[0183]

[0184] In the training process, L disengangle is used to train E and D. GAN-C After one update of E and D, L GAN-G is used to continue training E and D. (i) And L (i) is used to continue training E and D. (i) One adversarial training of C and G is used to improve the discrimination ability of C and the generation ability of G, so that the unsupervised intermediate samples generated by the exchange of the heterogeneous attribute factors gradually converge to the real samples.

[0185] Step three, unknown working condition factor reorganization and sample decoding generation

[0186] After the training of the working condition factor decoupling part, SIMDRN has the ability to store the information related to each attribute in the original sample into the factors in the latent variable in the encoding process, which are mutually exclusive and mutually exclusive. Suppose the information related to m working condition parameter attributes is stored in the factors in the latent variable. For a certain unknown working condition combination , the missing fault samples under this working condition can be obtained by encoding and extracting factors-reorganizing-decoding the existing samples under known working conditions.

[0187] First, according to the information in the shared attribute sample connection graph, select a sample set so that the ith attribute of the sample x (i) is the same as the ith attribute of the unknown working condition combination to be generated, that is Put all the samples in X into the trained encoder of SIMDRN to extract the latent variable:

[0188] z (i) = E(x (i) ), for i = 1, …, m

[0189] Where each latent variable is composed of m decoupled factors that control m attributes: Extract the factor that controls attribute i in the latent variable z (i) of each sample and reorganize it into a new latent variable:

[0190]

[0191] The reconstructed latent variable is input into the decoder for decoding to obtain a generated sample:

[0192]

[0193] Since each factor in the latent variable used for decoding is extracted from a known fault sample containing the target working condition attribute, the generated sample obtained by decoding satisfies the expected working condition combination Thus, the missing fault sample under the unknown working condition combination is reconstructed and generated.

[0194] The case analysis of the present application sets two cases. Among them, case 1 uses artificially generated simulation signal samples to preliminarily verify the factor decoupling ability and reconstruction generation ability of SIMDRN. Since real fault samples are complex and are disturbed by noise factors, the visualization is not good in terms of intuitive degree, therefore, the SIMDRN factor decoupling process and reconstruction generation process are visualized and analyzed using simulation signals with simple components and no noise influence. Case 2 uses the multi-condition, multi-fault condition aviation high-speed bearing data set disclosed by the Dynamic and Identification Research Group (DIRG) of the Department of Aerospace Engineering of a certain university, to verify the factor decoupling and generation performance of SIMDRN on complex real fault samples, and to compare with similar methods.

[0195] Case 1: Simulation data set

[0196] (11) Data set construction

[0197] In this case, a simulation signal sample is constructed by superimposing two sinusoidal signals with different frequencies, as follows:

[0198]

[0199] Wherein, the signal amplitudes A1 = 1, A2 = 1, the signal phases and are randomly sampled between 0 and 2π. The frequencies f1 and f2 are two attributes of the sample, respectively taking values in the sets {20, 40, 60} and {80, 100, 120}. The combination of f1 and f2 determines the frequency attribute of the sample.

[0200] The length of each sample is set to 1024 points, and the signal is sampled from the signal generator at a sampling frequency of 512 Hz, i.e. the sampling length of each sample is 2s. The number of simulation samples generated under each frequency combination condition is 200. The sample matrix is shown in Figure 3 The original signal samples and their spectra under nine frequency combination conditions are shown in Figures 4-5 ​as shown.

[0201] As can be seen from the spectrum, the frequency components contained in the superimposed sample are the combination of the frequency of each single frequency signal, so it can be directly judged whether the sample generated by the model meets the expected attribute requirements by observing the frequency components of the generated sample.

[0202] (12) Parameter setting

[0203] Since the simulation signal is relatively simple, the network structure constructed in this case is also a simple and lightweight network, and the specific structure and hyperparameters are shown in Table 1.

[0204] Table 1 Model structure and hyperparameters of case 1

[0205]

[0206] Among them, the vector output by Layer 6 of the encoding network is the latent variable containing the effective information of the original sample, with a dimension of 20. It is set that the first 10 dimensions are factors controlling the frequency f1 attribute of the generated sample, and the last 10 dimensions are factors controlling the frequency f2 attribute of the generated sample. The two factors are mutually exclusive, and the training target is to decouple the two factors. During the model training process, the batch size is 200 samples, and a total of 500 training iteration rounds are performed. The weights λ1, λ2, and λ3 of the three parts of the reconstruction loss during the working condition factor decoupling training process are 1, 0.25, and 0.5, respectively. The optimizer uses Adam, and the initial learning rate, learning rate decay coefficient, and other hyperparameters are set to the default values of the optimizer.

[0207] Since the simulation signal is simple, the discriminant network after the second exchange of factors is not set in this case to reduce the learning difficulty and the risk of overfitting of the model; at the same time, it is also convenient to more directly evaluate the implementation process and sample recombination generation performance of SIMDRN in attribute factor decoupling.

[0208] (13) SIMDRN model training and testing

[0209] The model is constructed according to the above parameters, and training is performed on the constructed dataset. During the training process, the changes of the original sample reconstruction loss, the same attribute factor one-time exchange reconstruction loss, and the different attribute factor two-time exchange reconstruction loss are as shown in the following figure. Figure 6

[0210] As can be seen from the loss curve, the three parts of the loss are effectively optimized and converged to near zero, indicating that SIMDRN can consider the three learning goals to enable the model to have the ability to effectively compress and represent the original sample and decouple the factors corresponding to different attributes.

[0211] ​The trained model is used to reconstruct the original sample, the same attribute factor once exchange, and the different attribute factor twice exchange process respectively to verify the learning effect of the model in these three aspects.

[0212] First, the samples with frequency combination (f1, f2) values of (20, 80), (60, 100), and (40, 120) are selected for reconstruction generation, and the generated samples are subjected to FFT processing to obtain the generated sample spectrum, as shown in Figure 7 .

[0213] The results show that the frequency components contained in the generated samples in the three cases are consistent with the expected and the same as the frequency components contained in the input original samples, and there is no other irrelevant frequency interference, verifying the model's ability to effectively compress and reconstruct the original samples.

[0214] Further, the samples with frequency combination (f1, f2) values of (40, 80) and (40, 100) are selected as the original samples, and after compression representation by the encoding network, the factor region corresponding to attribute f1 is exchanged, and the exchanged hidden variable is input into the decoding network for decoding to obtain the reconstruction generated sample after the same attribute factor once exchange and the frequency spectrum after FFT processing, as shown in Figure 8 .

[0215] Since the original samples have the same value in the attribute corresponding to the exchanged factor, according to the optimization objective, the generated sample by the model should be consistent with the original input sample. The results show that the frequency components contained in the decoded generated sample after exchange accurately fall in the position of the frequency components contained in the original input sample.

[0216] Similarly, the samples with frequency combination (f1, f2) values of (20, 100) and (60, 100) are selected for compression representation, the factor region corresponding to attribute f2 is exchanged, and the frequency spectrum after FFT transformation of the decoded reconstruction is obtained, as shown in Figure 9 .

[0217] The results show that the reconstructed sample is still consistent with the frequency components contained in the original sample. The results of the same attribute factor once exchange reconstruction generation for the two attributes show that the model can store the information related to a specific attribute in the corresponding hidden variable region at this time, i.e., it has the basic factor decoupling ability.

[0218] Finally, the samples with frequency combination (f1, f2) values of (60, 100) and (40, 120) are selected, once compression representation is performed, the factor corresponding to attribute f1 is exchanged and decoded for reconstruction generation; twice compression representation is performed, the factor corresponding to attribute f2 is exchanged and decoded for reconstruction generation. The visualization of the two exchange processes and the generated samples is as followsFigure 10 As shown in the figure.

[0219] For the two times of exchange process, the exchanged are the heterogeneous attribute factors, that is, the original samples take different values on the attributes corresponding to the exchange factors. And this also shows that in the training process, there is no supervision information to constrain the model, and after the exchange of the heterogeneous attribute factors, what attribute combination should the generated sample be reconstructed. The results of the above figure show that in the two times of exchange process of the heterogeneous attribute factors, the model can generate the sample with the correct frequency combination as expected. This further shows that the SIMDRN has a clear factor decoupling ability after training, and at the same time shows that the model can generate the sample with unknown attribute combination without real sample supervision.

[0220] (14) Decoupling representation visualization analysis

[0221] In theory, if the model can realize complete decoupling of the factors corresponding to different attributes, for a sample with the same attribute value, the factor corresponding to this attribute should fall in a cluster with close distance in the feature space; and for a sample with different attribute values, the factor corresponding to this attribute should fall in a region with far distance in the feature space. Therefore, the dimensionality reduction visualization analysis of the factors can directly verify the decoupling ability and effect of the model.

[0222] Select the sample with the frequency combination (f1, f2) taking the values of (20, 80), (40, 80), (60, 80), which have the same f2 value but different f1 values, and use the encoding network to compress the representation. Then, the f1 and f2 corresponding factor regions are extracted, reduced to 2 dimensions by the PCA algorithm, and the dimensionality reduction results are visualized as shown in the figure. Figure 11

[0223] Similarly, select the sample with the frequency combination (f1, f2) taking the values of (20, 80), (20, 100), (20, 120), which have the same f1 value but different f2 values, and use the encoding network to compress the representation. Then, the f1 and f2 corresponding factor regions are extracted, reduced to 2 dimensions by the PCA algorithm, and the dimensionality reduction results are visualized as shown in the figure. Figure 12

[0224] The two times of factor dimensionality reduction visualization results show that for the same attributes shared in the input sample, the corresponding factors are also mixed together in the feature space, indicating that the model considers that there is no obvious difference in the attribute of the input sample; and for the attributes with different values in the input sample, the corresponding factors are obviously clustered to form several regions with clear boundaries, indicating that the model considers that there is a significant difference in the attribute of the input sample. This visualization result further shows that the SIMDRN proposed in the present application has a clear factor decoupling representation ability.

[0225] ​​Case 2: Air high-speed bearing dataset

[0226] In this case, the dataset disclosed by DIRG is used to verify and analyze the SIMDRN model and the missing fault sample recombination generation method under unknown working condition combinations.

[0227] (21) DIRG air high-speed bearing dataset description

[0228] This dataset (hereinafter referred to as DIRG dataset in this case) is completed by the DIRG research group of the Department of Aerospace Engineering of a certain university of science and technology. The dataset is collected by a test bench, including variable rotating speed, radial load, fault bearing vibration data under fault degree, and also collects the environmental temperature in the test process. The main part of the test bench is composed of a high-speed spindle and a rotating shaft, and the basic structure is as shown in Figure 13 .

[0229] Bearing B1 and B3 jointly support the rotating shaft, and the radial load is applied to the position of B2 bearing through the action of the spring, and the size of the applied load is measured by a static force sensor. B1 is the fault bearing of the measured bearing, and two acceleration sensors A1 and A2 are fixed on the bearing seats of B1 and B2 respectively. By using the method of using Rockwell tools to make conical indentation on the bearing parts, the bearing is damaged to obtain the fault bearing. Bearings in different health states are replaced in B1 position to collect vibration data in different health states.

[0230] The dataset contains bearing data in 7 health states, as shown in Table 2.

[0231] Table 2: DIRG dataset health state label composition information

[0232]

[0233] For each health state of the measured bearing, the following test steps are experienced during the test process:

[0234] 1) Under no-load condition, rotate at 100Hz to confirm the correctness of the assembly;

[0235] 2) Apply static load in the order of 1000N, 1400N, and 1800N;

[0236] 3) Under each static load condition, increase the rotating speed of the rotating shaft from 0Hz to 500Hz at a step of 100Hz;

[0237] 4) When the rotating shaft reaches the set rotating speed and stabilizes, measure the acceleration and store it.

[0238] Due to the power limit of the equipment, the high rotating speed cannot be achieved under the condition of large radial load in the experiment. Therefore, in this case, only the data under the conditions of 100 Hz, 200 Hz, 300 Hz rotating speed are used. At the same time, since the bearing is usually working under certain load in the actual application scene, only the data under the conditions of 1000 N, 1400 N, 1800 N radial load are used in this case. In this case, rotating speed and load are two attributes that are concerned respectively, and different combinations of the two attributes form different working condition combinations. For each working condition combination under each health state, the original sample sampling frequency is 51.2 kHz, and the recording length is 10 s. The sample matrix under each health state is shown in FIG. 2. Figure 14

[0239] The original sample under each condition is divided into short samples with a length of 1024 points, and 40% of them are selected as training samples and added to the training set.

[0240] (22) Test parameter setting

[0241] Since the real fault sample is more complex than the simulation signal, a simple fully connected network may not be able to effectively learn the effective information in the data, so a lightweight convolutional structure is used to construct the model in this case. The structure and parameters of the encoding network and the decoding network are shown in Table 3.

[0242] Table 3 Model structure and hyperparameters of case 2

[0243]

[0244]

[0245] The output dimension of the encoder network is 100, that is, the dimension of the hidden variable obtained after the original sample is compressed is 100. The first 50 dimensions are set to be rotating speed factors representing rotating speed related information, and the last 50 dimensions are load factors representing load related information, and the two parts of the factors do not overlap and do not affect each other.

[0246] In this case, an intermediate sample discriminator is set. Except for the output layer, the intermediate sample discriminator network reconstructed after the one-time exchange of the two attributes factors has the same structure as the encoder network, and the output dimension of the output linear layer of the discriminator network is changed to 1.

[0247] ​In each health state, the samples in the four corners and the middle position of the sample matrix are set to be missing, i.e., (100Hz, 1000N), (100Hz, 1800N), (200Hz, 1400N), (300Hz, 1000N), and (300Hz, 1800N) are set as unknown working condition combination conditions, respectively. The SIMDRN is trained by using the fault samples under the known working condition combination conditions in the sample matrix. The trained SIMDRN generator network is used to recombine and generate the missing samples under unknown working condition combinations.

[0248] In the training process, the sample batch size used for performing a parameter update is 100, and the training iteration rounds are 500. The optimizer uses Adam, and the initial learning rate, learning rate decay coefficient, and other hyperparameters are set to the default values of the optimizer.

[0249] (23) Visualization analysis of recombination and generation results of missing fault samples

[0250] Taking the health state H1 (inner ring fault, 450 microns) as an example, the missing fault samples under each unknown working condition combination are recombined and generated, a generated sample is randomly selected, the original time domain waveform and the frequency spectrum after FFT transformation of the sample are visualized, and are compared with the real sample under the corresponding condition as shown in Figures 15-19 .

[0251] The recombination and generation results show that the generated sample and the real sample have high consistency in the time domain and the frequency domain. The fault samples under different working condition combinations have different frequency compositions, SIMDRN can learn the influence of different working condition attributes on the samples and compress the representation in the form of decoupling factors, and then generate unknown working condition combinations. For example, the samples under the conditions of (300Hz, 1000N) and (300Hz, 1800N) have high frequency components in the low frequency part, which do not appear in the fault samples under other working condition combination conditions. SIMDRN can still accurately generate them, which shows that the model is not a simple repetition or modification of the existing samples.

[0252] Single sample visualization can show the consistency of the generated sample itself with the real sample, but cannot describe the ability and quality of the generated samples from the data distribution level. Therefore, further dimensionality reduction visualization analysis is performed on the generated samples and the real samples. Taking the missing condition of the fault sample under the condition of (200Hz, 1400N) as an example, the missing sample is recombined and generated; the generated fault sample and the real fault sample are reduced to the same two-dimensional space by using the t-SNE algorithm, and are visualized in the form of a scatter plot as shown in Figure 20 .

[0253] Figure 20Different scatter points represent different working condition combination conditions of fault samples. Except that (200Hz, 1400N) is a generated sample, the rest are real samples under different conditions. As can be seen, although all the samples in the figure come from the same fault mode, due to different working condition combinations, there are differences between the distributions of the samples in the space. At the same time, the generated fault sample does not coincide with the real sample under the condition of other known working condition combinations, which further indicates that the model is not repeating the existing samples in the training set, but generating new fault samples under unknown working condition combinations that are not included in the training set according to the set factor combination.

[0254] In the scatter plot drawn by using all samples for common dimension reduction, samples with the same speed and samples with the same load are respectively boxed and labeled with free curves as shown in Figure 21

[0255] According to the view of manifold learning, assuming that the original fault samples are low-dimensional manifolds falling in a high-dimensional data space, after reducing the original samples to low dimensions, the sample points with continuous changes in a certain attribute and the same in all other attributes should fall in a continuous low-dimensional space, and the trend of attribute change should be consistent with the direction of sample point distribution change. In the above figure, the samples with a speed of 100Hz and 300Hz form a region, and when the load changes from 1000N to 1400N to 1800N, the sample points basically change continuously in the same direction; similarly, in the region formed by the samples with a load of 1000N and 1800N, the same characteristics are shown. This also proves the rationality of the manifold learning assumption. Based on this assumption, when the speed is 200Hz, the generated fault sample with a load of 1400N falls between the real fault samples with a load of 1000N and 1800N; when the load is 1400N, the generated sample with a speed of 200Hz also falls between the real fault samples with a speed of 100Hz and 300Hz. This shows that from the perspective of data distribution, the fault sample generated by SIMDRN can basically keep consistent with the real situation.

[0256] The preferred embodiments of the present application are described above with reference to the accompanying drawings, and the scope of the right of the present application is not limited thereby. Any modification, equivalent replacement and improvement made by those skilled in the art without departing from the scope and essence of the present application shall be within the scope of the right of the present application.​

Claims

1. A method for reorganizing unknown working condition combined fault samples based on multi-attribute factor decoupling, characterized in that, The method comprises the following steps: acquiring a fault sample set of known working condition combinations, and constructing a shared attribute sample joint graph by using the working condition combination attributes of each fault sample in the fault sample set; constructing a deep generative model of an encoding-decoding structure, and training the deep generative model of the encoding-decoding structure by using the shared working condition attribute relationships between the fault samples in the shared attribute sample joint graph, to obtain a deep generative model of an encoding-decoding structure having a function of decoupling hidden variable factors corresponding to multiple working condition attributes, which comprises: sequentially performing original sample compression reconstruction, first attribute sample factor exchange reconstruction and second attribute sample factor exchange reconstruction training on the deep generative model of the encoding-decoding structure by using the shared working condition attribute relationships between the fault samples in the shared attribute sample joint graph, to obtain the deep generative model of the encoding-decoding structure having the function of decoupling the hidden variable factors corresponding to the multiple working condition attributes; acquiring an unknown working condition combination to be generated, and selecting a fault sample set of a known working condition combination having the same working condition attributes as the unknown working condition combination to be generated according to the working condition attributes of the unknown working condition combination to be generated; inputting the fault sample set of the known working condition combination having the same working condition attributes into the deep generative model of the encoding-decoding structure having the function of decoupling the hidden variable factors corresponding to the multiple working condition attributes, to obtain a reorganized generated sample of a missing fault sample under the unknown working condition combination; wherein the first attribute sample factor exchange reconstruction training on the deep generative model of the encoding-decoding structure comprises: Original sample encoding process: , ; Hidden variable factor interchanging process: ; Interchangeable latent variable decoding process: ;​ Computing the co-property sample factor one-exchange reconstruction loss: ; wherein, , is a latent variable encoded from a sample , is a latent variable , is a new latent variable obtained after swapping the j-th factor; , is a generated sample decoded from the latent variable ​​ the second attribute sample factor exchange reconstruction training on the deep generative model of the encoding-decoding structure comprises: Selecting a pair of fault samples and input each fault sample in the pair of fault samples into the encoder respectively to obtain a latent variable: ； On latent variables and Interchange the factors corresponding to attribute j, resulting in an interchanged latent variable: ； Swap latent variables Input decoder, to decode the generated samples after one swap: ； Again, the generated sample after the first exchange and is fed into the encoder, resulting in the secondly encoded latent variable: ； Hidden variables for quadratic encodings and The factors corresponding to attribute j are permuted again, resulting in quadratic permuted hidden variables: ； twice the exchange of hidden variables input decoder, and decode the generated sample after the twice exchange ； using a mean square error to measure the deviation between the pair of secondly exchanged generated samples and the pair of original samples as a second attribute sample factor exchange reconstruction loss: 。 2. The method of claim 1, wherein, the construction of the shared attribute sample joint graph by using the multiple working condition attributes of each fault sample in the fault sample set comprises: acquiring the fault mode and working condition combination attributes of each fault sample in the fault sample set; dividing the fault sample set according to the fault mode to obtain a plurality of first fault sample sets; dividing the first fault sample sets according to the working condition combination attributes to obtain a plurality of second fault sample sets; constructing a shared attribute sample joint graph by using the multiple second fault sample sets of each fault sample in the fault sample set.

3. The method of claim 2, wherein, the construction of the shared attribute sample joint graph by using the multiple second fault sample sets of each fault sample in the fault sample set comprises: taking each fault sample in the fault sample set as a node, taking the shared working condition attribute relationship between two fault samples as an edge, and taking the attribute label corresponding to the shared working condition attribute as the value of the edge, to construct a shared attribute sample joint graph.

4. The method of claim 1, wherein, the original sample compression reconstruction training on the deep generative model of the encoding-decoding structure comprises: The original sample compression reconstruction loss utilizes the shared working condition attribute relationship between each fault sample in the shared attribute sample connection graph, and calculates the mean square error between each original fault sample and the reconstructed sample obtained by encoding-decoding as the original sample compression reconstruction loss ; ； wherein, is the original sample compression reconstruction loss, is the desired function, is the compressed reconstructed sample, is the original sample.

5. The method of claim 1, wherein, The process of inputting the set of fault samples from known operating condition combinations with the same operating condition attributes into a deep generative model with an encoder-decoder structure that decouples latent variable factors corresponding to multiple operating condition attributes, to obtain reconstructed fault samples from unknown operating condition combinations includes: Using the encoder network in the deep generative model with the function of decoupling latent variable factors corresponding to multiple working condition attributes, the fault samples of known working condition combinations are encoded into latent variables related to working condition factors. Through the cross-interchange between latent variables, latent variables containing information of unknown working condition combinations are obtained. By utilizing the decoder network in the deep generative model with the function of decoupling latent variable factors corresponding to multiple operating condition attributes, the latent variables containing unknown operating condition combination information are decoded into fault samples, thereby realizing the reorganization and generation of missing fault samples under unknown operating condition combinations.

6. The method of claim 5, wherein, The encoder network in the deep generative model, which utilizes an encoder-decoder structure with decoupling functionality for latent variable factors corresponding to multiple operating condition attributes, encodes fault samples of known operating condition combinations into latent variables related to operating condition factors. Through cross-interchange between latent variables, latent variables containing information about unknown operating condition combinations are obtained, including: All samples from the fault sample set of known operating condition combinations with the same operating condition attributes as the stated operating condition are fed into the encoder network to extract latent variables: ； where each latent variable is composed of m decoupled factors controlling m attributes: ; extract the factor controlling attribute i in each sample's latent variable and reorganize it into a new latent variable as the latent variable containing unknown working condition combination information: 。 7. The method of claim 6, wherein, The step of using the decoder network in the deep generative model with an encoder-decoder structure that decouples latent variable factors corresponding to multiple operating condition attributes to decode the latent variables containing unknown operating condition combination information into fault samples, thereby realizing the reconstructive generation of missing fault samples under unknown operating condition combinations, includes: The input decoder network, which contains information about unknown operating condition combinations, is used to decode the missing fault samples under the unknown operating condition combinations, and the samples are reconstructed to generate samples: 。

Images