Sensor fault diagnosis method based on single-domain generalization under uncertainty guidance adversarial enhancement domain

Abstract

The invention discloses a single-domain generalization sensor fault diagnosis method based on an uncertainty guidance adversarial enhancement domain. The method comprises the following steps: obtaining operation data of a target sensor; inputting the operation data into a trained sensor fault diagnosis model, and outputting a fault category corresponding to the operation data; the training process of the sensor fault diagnosis model comprises the following steps: in a forward propagation process, generating an adversarial source domain and an imaginary domain based on a single source domain of a sensor; in the back propagation process, uncertainty estimation is carried out on the adversarial source domain and the imaginary domain, and an enhanced domain is generated; and training a learning class decision boundary of the sensor fault diagnosis model through the enhancement domain. According to the method, training of the sensor fault diagnosis model can be realized only based on a single source domain, strict requirements of multiple source domains in the training process of the fault diagnosis model are eliminated, and known faults and unknown faults can be identified under unknown working conditions; according to the method, uncertainty estimation is introduced, construction of an enhancement domain is supervised, and effectiveness of the method is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of single-domain generalized fault diagnosis, and more specifically to a sensor fault diagnosis method based on single-domain generalization in an uncertainty-guided adversarial enhancement domain. Background Art

[0002] As modern industrial systems become increasingly complex, the scale of sensor networks has gradually expanded, bringing maintenance challenges. Especially under the influence of factors such as loss, aging and harsh environment, it is easy to cause sensor failures and even production accidents. Although manual calibration can extend the life of sensors and improve safety, this method is costly and time-consuming in the face of the increasing number of sensors. Therefore, the development of advanced sensor fault diagnosis technology, especially fault diagnosis methods that can adapt to multiple working conditions, is of vital importance to ensure the stable operation of industrial systems and improve production efficiency.

[0003] In the past decade, with the rapid development of artificial intelligence technology, deep learning-based fault diagnosis (DLFD) methods have attracted much attention due to their fast response speed, high diagnostic accuracy, and low requirement for expert knowledge. However, most of the current DLFD methods assume that the training phase and the test phase (represented as the source domain and the target domain, respectively) have the same distribution. However, in actual industrial processes, this strict assumption is difficult to hold. Due to changes in operating conditions and the influence of external interference, the distribution of source domain and target domain data in industrial processes is usually quite different, resulting in a decrease in the performance of DLFD methods.

[0004] In order to solve the above problems, some scholars have introduced domain adaptation (DA) technology in recent years, and the domain adaptation-based fault diagnosis (DAFD) method to solve the cross-domain fault diagnosis problem under different working conditions. This method has performed well in many theoretical research processes, but it still has limitations when dealing with actual cross-domain fault diagnosis tasks. First, DAFD requires the participation of the target domain in the model training process to obtain the distribution information of the target domain. Second, the actual cross-domain fault diagnosis task is usually carried out under new working conditions. Since it is impossible to obtain unlabeled target domain data during the model training stage, it is difficult to identify unknown faults.

[0005] In order to better solve the above-mentioned cross-domain fault diagnosis problem, domain generalization (DG) has gradually received attention in recent years. The domain generalization-based fault diagnosis (DGFD) method gets rid of the dependence on target domain data. By generalizing the diagnostic model to unknown working conditions, it can effectively deal with the problem of being unable to obtain unlabeled target domain data during the model training phase and making it difficult to identify unknown faults. Although the DGFD method does not rely on prior knowledge of the target domain, it needs to be trained on multiple source domains. However, in actual industrial applications, it is difficult to obtain data under multiple working conditions, which greatly reduces the application scope of the traditional DGFD method. It can be seen that when the distribution of training samples is single, single domain generalization (SDG) is more practical. SDG technology alleviates the constraints of too many restrictions on multiple source domains, usually by increasing samples through data enhancement or expansion, thereby meeting the needs of data distribution diversity.

[0006] With the development of SDG technology, single domain generalization-based fault diagnosis (SDGFD) methods have begun to attract attention. Zhao et al. proposed a single domain generalization network (AMINet) guided by adversarial mutual information, which used a domain generation module to generate pseudo target samples with new distributions and realized intelligent fault diagnosis of unknown working conditions through adversarial mutual information algorithm. Wang et al. proposed a multi-scale style generative adversarial contrastive single domain generalization network (MSG-CAN), which generated diversified samples through a multi-scale style learning strategy and introduced contrastive learning in the adversarial network, which facilitated the learning of domain-invariant representations. However, the above two methods still have certain limitations: (1) The above two methods do not add generation constraints in the generation module, which may lead to the generation of samples that are irrelevant to semantic information; (2) The situation of new fault modes under unknown working conditions is not discussed in the experimental part; (3) Before the task model performs domain-invariant learning, an uncertainty estimate is not performed on the generated samples, which easily leads to the task model learning invalid generalization information during the training process. It can be seen that the relevant research on the SDGFD method has not been fully carried out and still needs further exploration.

[0007] Therefore, how to eliminate the stringent requirements of multiple source domains in the training process of sensor fault diagnosis models, and ensure that the sensor fault diagnosis model can identify known faults and unknown faults under unknown working conditions, while introducing uncertainty generalization estimation to supervise the construction of enhanced domains and ensure its effectiveness, is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0008] In view of the above problems, the present invention provides a sensor fault diagnosis method based on single-domain generalization in an uncertainty-guided adversarial enhancement domain, so as to at least solve some of the technical problems mentioned in the above background technology.

[0009] In order to achieve the above object, the present invention adopts the following technical solution:

[0010] The present invention provides a sensor fault diagnosis method based on single-domain generalization in an uncertainty-guided countermeasure enhancement domain, comprising the following steps:

[0011] Acquire operating data of a target sensor; input the operating data into a trained sensor fault diagnosis model, and output a fault category corresponding to the operating data;

[0012] The training process of the sensor fault diagnosis model includes a forward propagation process and a back propagation process;

[0013] In the forward propagation process, an adversarial source domain and a fictitious domain are generated based on a single source domain of a sensor;

[0014] In the back-propagation process, uncertainty estimation is performed on the adversarial source domain and the imaginary domain to generate an enhanced domain; and the sensor fault diagnosis model is trained by using the enhanced domain to learn a class decision boundary.

[0015] Furthermore, the sensor fault diagnosis model is constructed by an uncertainty-guided adversarial enhancement domain network;

[0016] The uncertainty guided adversarial enhancement domain network includes a generation module, an out-of-domain auxiliary module and an uncertainty domain adversarial network;

[0017] The out-of-domain auxiliary module is constructed by a multi-scale random convolutional network;

[0018] The uncertainty domain adversarial network includes a domain feature extractor, a domain identifier and a diagnosis task module.

[0019] Furthermore, the forward propagation process specifically includes:

[0020] Acquire a single source domain of a sensor; the single source domain includes a source sample and a fault category label corresponding to the source sample; the source sample includes a plurality of historical operation data of the sensor;

[0021] In the generation module, using Wasserstein distance as a constraint condition, an adversarial sample consistent with the semantic information of the source sample is generated, and an adversarial source domain is constructed according to the adversarial sample;

[0022] In the out-of-domain auxiliary module, the source samples are projected into the drift domain as a relaxation process to generate pseudo samples; and a fictitious domain is constructed based on the pseudo samples as a learning object of the generation module.

[0023] Furthermore, the loss function corresponding to the constraint condition is expressed as:

[0024]

[0025] Among them, L cons represents the loss function of the constraint condition; θ G According to L cons Optimize the parameters of WAE; WAE stands for Wasserstein distance-based autoencoder; D s represents a single source domain; D s+ represents the adversarial source domain; d(·) represents the similarity estimation function; ρ represents the maximum domain difference; E(·) represents the calculation of mathematical expectation; W d represents Wasserstein distance; G represents generation module; En represents encoder; x i represents the i-th historical running data of the input; De represents the decoder; z i represents the i-th latent space feature; p(z) represents the latent space feature z i The prior distribution of .

[0026] Furthermore, the loss function corresponding to the relaxation process is expressed as:

[0027]

[0028] Among them, L rela represents the loss function of the relaxation process; θ D According to L rela Optimize the parameters of multi-scale random convolutional networks; D s represents a single source domain; D f represents the fictitious domain; d(·) represents the similarity estimation function; ρ represents the maximum domain difference; D(·) represents the calculation of variance; x s represents the source sample; represents the metric distribution function; p(z) represents the latent space feature z i The prior distribution of .

[0029] Furthermore, the back propagation process specifically includes:

[0030] Extracting features from the adversarial source domain and the imaginary domain by the domain feature extractor to obtain a high-level distribution representation of the domain;

[0031] Inputting the high-level distribution representation into the domain identifier and outputting a confidence logarithm; generating a confidence estimate value by applying the confidence logarithm to an S-curve;

[0032] Input the high-level distribution representation into the diagnostic task module and output the predicted logarithm; apply the predicted logarithm to the Softmax function to obtain the class prediction probability;

[0033] Based on the confidence estimate and the class prediction probability, uncertainty estimation is performed on the adversarial source domain and the imaginary domain to generate enhanced samples; and an enhanced domain is constructed according to the enhanced samples;

[0034] The diagnostic task module is trained through the enhanced domain so that the diagnostic task module learns the class decision boundary.

[0035] Furthermore, the step of performing uncertainty estimation on the adversarial source domain and the imaginary domain based on the confidence estimate and the class prediction probability to generate enhanced samples; and constructing an enhanced domain according to the enhanced samples specifically includes:

[0036] Taking the correct fault classification label as a hint, we adjust between the confidence estimate c and the class prediction probability p by interpolation to obtain the optimized class prediction probability; specifically, it is expressed as:

[0037] p i ′=c·p i +(1-c)y i

[0038] Among them, p i Represents x i The corresponding class prediction probability; x i represents the i-th historical running data input; p i ′ indicates that p i The optimized class prediction probability; c represents the confidence estimate; y i Indicates the value of the i-th element of the fault category label in the one-hot encoding, i.e. x i The corresponding fault category label;

[0039] Based on the optimized class prediction probability, the diagnostic task loss of the diagnostic task module is calculated; specifically, it is expressed as:

[0040]

[0041] Among them, L task represents the diagnostic task loss; M represents the total amount of input historical operation data;

[0042] Based on the loss function of the constraint condition, the loss function of the relaxation process, and the diagnostic task loss, the total task loss function of the uncertainty-guided adversarial enhancement domain network is obtained; specifically expressed as:

[0043] L UAAD =L task (θ T ;x s )-λL cons (θ G ; z)+βL rela (θ D ;x s )

[0044] Among them, L UAAD represents the total task loss function of the uncertainty-guided adversarial enhancement domain network; θ T represents the parameters of the optimized diagnosis task module; z represents the latent space feature; λ represents the balance L cons Hyperparameters; β represents the balance L rela Hyperparameters of

[0045] Based on the total task loss function L UAAD , forming enhanced samples and constructing enhanced domains; specifically expressed as:

[0046]

[0047] Among them, θ G,D,T Represents θ G ,θ D and θ T ; represents the source sample after t enhancement iterations; z t represents the latent space features after t enhancement iterations; represents the adversarial sample after t enhancement iterations; Indicates that Input into the uncertainty domain adversarial network for enhanced iteration; η represents the learning rate of gradient ascent; represents the enhanced sample after t enhancement iterations; Indicates based on Constructed enhancement domain.

[0048] Furthermore, the step of training the diagnostic task module through the enhanced domain so that the diagnostic task module learns the class decision boundary specifically includes:

[0049] The diagnostic task module is enabled to learn sample features of each fault category in the single source domain, and the average feature of each fault category is used as the feature center;

[0050] The samples in the single source domain and the enhanced domain are input into the diagnostic task module for classification, and the distance between each correctly classified sample feature and the average feature under the same fault category is calculated; specifically expressed as:

[0051]

[0052] in, represents the distance between the correctly classified sample features and the average features under the kth fault category; K represents a total of K fault categories; L CE represents the cross entropy function; represents the sample features of the kth fault category in the adversarial source domain; express Fault category label of J k represents the feature center under the kth fault category; C represents the classifier;

[0053] In the diagnostic task module, the distribution difference of the domain sample space is measured by the conditional maximum average deviation to optimize the distance between each correctly classified sample feature and the average feature under the same fault category; specifically, it is expressed as:

[0054]

[0055] Among them, L cmmd represents the loss function of the optimized distance; i represents the i-th historical running data; φ represents the kernel density function; represents the sample features of the kth fault category in the enhanced domain;

[0056] A corresponding class decision boundary is constructed for each known fault category so that the diagnostic task module can identify known faults and unknown faults; specifically, it is expressed as:

[0057]

[0058] in, represents the class decision boundary corresponding to the k-th fault category.

[0059] Furthermore, it also includes: optimizing the class decision boundary of each fault category, which is specifically expressed as:

[0060]

[0061]

[0062] Among them, L CDB represents the loss function that optimizes the class decision boundary; θ k represents the variables that control the effects of the two constraints; represents the decision boundary of the kth fault class.

[0063] Furthermore, it also includes optimizing the uncertainty-guided countermeasure enhancement domain network; specifically:

[0064] The network optimization problem is expressed as:

[0065]

[0066]

[0067] in, Indicates the optimized parameters of the generated module; represents the optimized parameters of the out-of-domain auxiliary module; Represents the optimized parameters of the diagnostic task module; L cons represents the loss function of the constraint condition; θ G According to L cons Optimize the parameters of WAE; L rela represents the loss function of the relaxation process; θ D According to L rela Optimize the parameters of multi-scale random convolutional networks; L task represents the diagnosis task loss; θ T According to L task Optimize the parameters of the diagnostic task module; represents the optimized parameters of the domain discriminator; represents the optimized parameters of the domain feature extractor; represents the parameters of the optimized domain discriminator; θ F represents the parameters of the optimized domain feature extractor; E(·) represents the calculation of mathematical expectation; θ C represents the optimized classifier;

[0068] The Adam optimizer is used to solve the network optimization problem and update the network parameters in each training round. The specific expression is:

[0069]

[0070]

[0071]

[0072] Among them, μ represents the learning rate; represents the generation module parameters in the e-th iteration update process; θ G e+1 represents the parameters of the generated module in the e+1th iteration update process after e iteration updates; represents the out-of-domain auxiliary module parameters in the e-th iteration update process; represents the out-of-domain auxiliary module parameters in the e+1th iterative update process after e iterative updates; Represents the diagnostic task module parameters in the e-th iteration update process; represents the diagnostic task module parameters in the e+1th iteration update process after e iteration updates; L CDB represents the loss function that optimizes the class decision boundary; represents the class decision boundary corresponding to the kth fault category in the eth iterative update process; represents the class decision boundary corresponding to the kth fault category in the e+1th iterative update process after e iterations; λ represents the balance L cons Hyperparameters; β represents the balance L rela The hyperparameter υ represents the weight parameter for measuring the decision boundary during the training process of the uncertainty domain adversarial network.

[0073] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a sensor fault diagnosis method based on single-domain generalization under uncertainty-guided countermeasure enhancement domain, which has the following beneficial effects:

[0074] The present invention can train the sensor fault diagnosis model based on a single source domain, which eliminates the stringent requirement of multiple source domains in the training process of the fault diagnosis model and ensures that the sensor fault diagnosis model can identify known faults and unknown faults under unknown working conditions. In addition, the present invention introduces uncertainty estimation to supervise the construction of the enhanced domain to ensure its effectiveness.

[0075] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0076] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0078] Figure 1 A schematic flow chart of a sensor fault diagnosis method based on single-domain generalization in an uncertainty-guided countermeasure enhancement domain provided in an embodiment of the present invention.

[0079] Figure 2A schematic diagram of the uncertainty-guided adversarial enhancement domain network framework provided in an embodiment of the present invention.

[0080] Figure 3 A schematic diagram of the process flow and operating position of a nickel flash furnace system provided in an embodiment of the present invention.

[0081] Figure 4 A schematic diagram of a sensor failure mode provided by an embodiment of the present invention.

[0082] Figure 5 A schematic diagram of the diagnostic results of the comparison method provided in an embodiment of the present invention.

[0083] Figure 6 A schematic diagram of learning features of visualization task T3 provided in an embodiment of the present invention.

[0084] Figure 7 A schematic diagram of enhanced domain feature distribution under unknown working conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] See also Figure 1 As shown, the embodiment of the present invention discloses a sensor fault diagnosis method based on single-domain generalization in an uncertainty-guided countermeasure enhancement domain, comprising the following steps:

[0087] Obtain the operating data of the target sensor; input the operating data into the trained sensor fault diagnosis model, and output the fault category corresponding to the operating data;

[0088] The training process of the sensor fault diagnosis model includes the forward propagation process and the back propagation process;

[0089] In the forward propagation process, adversarial source domain and imaginary domain are generated based on the single source domain of the sensor;

[0090] During the back-propagation process, uncertainty estimation is performed on the adversarial source domain and the imaginary domain to generate an enhanced domain; the sensor fault diagnosis model is trained through the enhanced domain to learn the class decision boundary.

[0091] This method can train the sensor fault diagnosis model based on a single source domain, which eliminates the strict requirements of multiple source domains in the training process of the fault diagnosis model and ensures that the sensor fault diagnosis model can identify known faults and unknown faults under unknown working conditions. In addition, this method introduces uncertainty estimation and supervises the construction of the enhanced domain to ensure its effectiveness.

[0092] Next, the training method of the above sensor fault diagnosis model is described in detail.

[0093] 1. Construction of sensor fault diagnosis model

[0094] See also Figure 2 As shown in the figure, the sensor fault diagnosis model is constructed by the uncertainty-guided adversarial enhancement domain network (UAAD); the uncertainty-guided adversarial enhancement domain network (UAAD) includes a generation module G, an out-of-domain auxiliary module D and an uncertainty domain adversarial network (UADNet); where:

[0095] The above-mentioned out-of-domain auxiliary module D is constructed by a multi-scale random convolutional network (MRCNN), which can construct new domains with similar global information but different local features;

[0096] The above uncertainty domain adversarial network (UADNet) includes domain feature extractor F, domain discriminator D ξ And diagnostic task module T.

[0097] 2. Training of sensor fault diagnosis model

[0098] 1. In the forward propagation process

[0099] (1) Obtain the single source domain D of the sensor s ; The single source domain D s Include source sample X s , and source sample X s Corresponding fault category label Y s ; Source sample X s contains multiple historical operation data of sensors; based on this, a single source domain D s Expressed as

[0100] (2) In the generation module G, the Wasserstein distance is used as a constraint to generate s Adversarial example X with consistent semantic information s+ , and according to the adversarial sample X s+ Construct adversarial source domain D s+ ;

[0101] Specifically, Wasserstein distance, also known as Earth Mover's distance (EMD), is a method for measuring the distance between variable distributions. The Wasserstein distance-based autoencoder (WAE) is inspired by the generative adversarial network (GAN) and introduces a latent space discriminator in the process of generating adversarial samples, so that it can retain semantic information while continuously optimizing the distribution. Generating adversarial samples through the Wasserstein autoencoder can improve the diversity of samples.

[0102] In this embodiment of the present invention, in order to ensure that the adversarial sample x s+ Preserve the semantic information of the source domain and the source sample x s Form sample pairs and establish constraints; the loss function corresponding to the constraints is expressed as:

[0103]

[0104] Among them, L cons represents the loss function of the constraint condition; θ G According to L cons Optimize the parameters of WAE; WAE stands for Wasserstein distance-based autoencoder; D s represents a single source domain; D s+ represents the adversarial source domain; d(·) represents the similarity estimation function; ρ represents the maximum domain difference; E(·) represents the calculation of mathematical expectation; W d represents Wasserstein distance; G represents generation module; En represents encoder; x i represents the i-th historical running data of the input; De represents the decoder; z i represents the i-th latent space feature; p(z) represents the latent space feature z i The prior distribution of

[0105] Constraint loss function L cons The establishment of provides a prerequisite for the subsequent construction of the enhanced domain, which effectively avoids the generation of adversarial samples that are irrelevant to the semantic information of the source samples during the training process;

[0106] (3) In the out-of-domain auxiliary module D, the source sample X s Generate pseudo samples by projecting the source distribution into a drift domain with arbitrary distribution; the drift domain refers to any new domain space that is allowed to change, used to generate new samples; then, construct a fictitious domain D based on the pseudo samples f As the learning object of the generation module G; the embodiment of the present invention defines this process as a relaxation process, specifically refers to the process in which the model parameters or variable distribution gradually approach the optimal solution with iteration; based on this, the loss function corresponding to the relaxation process is established, which is expressed as:

[0107]

[0108] Among them, L rela represents the loss function of the relaxation process; θ D According to L rela Optimize the parameters of multi-scale random convolutional networks; D s represents a single source domain; D f represents the fictitious domain; d(·) represents the similarity estimation function; ρ represents the maximum domain difference; D(·) represents the calculation of variance; x s represents the source sample; represents the metric distribution function; p(z) represents the latent space feature z i The relaxation process is another prerequisite for constructing the enhancement domain.

[0109] 2. In the back propagation process

[0110] (1) Through the domain feature extractor F, the adversarial source domain D s+ and the imaginary domain D f Perform feature extraction to obtain a high-level distribution representation of the domain: F(x) = x h ;

[0111] (2) Represent the high-level distribution x h Input to the domain discriminator D ξ In the output, the confidence logarithm is passed through an s-shaped curve to generate a confidence estimate c. When the network is confident that it can make a correct prediction for a given input, c should be close to 1. On the contrary, if the network does not believe that it can make a correct prediction, then c should be close to 0.

[0112] (3) Represent the high-level distribution x h Input into the diagnostic task module T, output the predicted logarithm; pass the predicted logarithm through the Softmax function to obtain the class prediction probability p;

[0113] (4) Based on the confidence estimate c and class prediction probability p, the adversarial source domain D s+ and the imaginary domain D f Estimate uncertainty and generate enhanced samples; construct the enhanced domain D based on the enhanced samples + ; Specifically include:

[0114] ① In order to improve the confidence of the network, the correct answer, that is, the correct fault classification label y i As a reminder, we interpolate between the confidence estimate c and the class prediction probability p to obtain the optimized class prediction probability; specifically, it is expressed as:

[0115] p i′=c·p i +(1-c)y i

[0116] Among them, p i Represents x i The corresponding class prediction probability; x i represents the i-th historical running data input; p i ′ indicates that p i The optimized class prediction probability; c represents the confidence estimate; y i Indicates the value of the i-th element of the fault category label in the one-hot encoding, i.e. x i The corresponding fault category label;

[0117] ② Based on the optimized class prediction probability p i ′, calculate the diagnostic task loss of the diagnostic task module T; specifically expressed as:

[0118]

[0119] Among them, L task represents the diagnostic task loss; M represents the total amount of input historical operation data;

[0120] In order to prevent the network from always choosing c = 0, a penalty function L is added to the diagnosis task loss c :

[0121] L c = -log(c)

[0122] Then we get the total loss L of the uncertainty domain adversarial network (UADNet):

[0123] L=L task +γL c

[0124] Among them, γ represents the weighted hyperparameter of trust loss;

[0125] ③ Loss function L based on constraints cons , the loss function L of the relaxation process rela , and the diagnosis task loss L task , the total task loss function of the uncertainty-guided adversarial enhancement domain network (UAAD) is obtained; specifically expressed as:

[0126] L UAAD =L task (θ T ;x s )-λL cons (θ G ; z)+βL rela (θ D ;xs )

[0127] Among them, L UAAD represents the total task loss function of the uncertainty-guided adversarial enhancement domain network; θ T represents the parameters of the optimized diagnosis task module; z represents the latent space feature; λ represents the balance L cons Hyperparameters; β represents the balance L rela Hyperparameters of

[0128] ④ Based on the total task loss function L UAAD , forming enhanced samples and constructing enhanced domain D + ; expressed as:

[0129]

[0130] Among them, θ G,D,T Represents θ G ,θ D and θ T ; represents the source sample after t enhancement iterations; z t represents the latent space features after t enhancement iterations; represents the adversarial sample after t enhancement iterations; Indicates that Input into the uncertainty domain adversarial network for enhanced iteration; η represents the learning rate of gradient ascent; represents the enhanced sample after t enhancement iterations; Indicates based on Constructed enhancement domain;

[0131] In this process, under the uncertainty domain adversarial network (UADNet), the constraints and relaxation process are balanced through uncertainty estimation, so that the adversarial samples can be effectively expanded outside the source domain to form a reasonable enhancement domain, which helps to improve the generalization ability of the model under training in the source domain and the enhancement domain.

[0132] (5) By enhancing the domain D + The diagnostic task module T is trained so that the diagnostic task module T learns the class decision boundary; specifically, it includes:

[0133] ① Make the diagnostic task module T learn a single source domain D s The sample features of each fault category in And the average feature under each fault category is taken as the feature center, that is,

[0134] ② The single source domain D s and enhanced domain D +The samples in are input into the diagnosis task module T for classification, and the distance between each correctly classified sample feature and the average feature under the same fault category is calculated; specifically expressed as:

[0135]

[0136] in, represents the distance between the correctly classified sample features and the average features under the kth fault category; K represents a total of K fault categories; L CE represents the cross entropy function; represents the sample features of the kth fault category in the adversarial source domain; express Fault category label of J k represents the feature center under the kth fault category; C represents the classifier;

[0137] ③ In the diagnostic task module T, the distribution difference of the domain sample space is measured by the conditional maximum average deviation to achieve the distance between each correctly classified sample feature and the average feature under the same fault category. Optimize; specifically:

[0138]

[0139] Among them, L cmmd represents the loss function of the optimized distance; i represents the i-th historical running data; φ represents the kernel density function; represents the sample features of the kth fault category in the enhanced domain;

[0140] ④ Construct a corresponding class decision boundary for each known fault category so that the diagnostic task module can identify known faults and unknown faults; specifically expressed as:

[0141]

[0142] in, represents the class decision boundary corresponding to the k-th fault category.

[0143] ⑤ Although most known sensor samples will be correctly classified into the class decision boundary, some free samples and unknown fault samples still have the risk of being misclassified into other known classes; this means that the empirical risk and generalization risk increase. In order to reduce these risks, the class decision boundary of each fault category is optimized in the embodiment of the present invention, which is specifically expressed as:

[0144]

[0145]

[0146] Among them, L CDB represents the loss function that optimizes the class decision boundary; θ k represents the variables that control the effects of the two constraints; represents the decision boundary of the kth fault category;

[0147] By learning class decision boundaries, specific decision boundaries are divided for different classes; these class decision boundaries cover most of the known sample patterns and can accurately distinguish unknown fault modes.

[0148] 3. Optimize the Uncertainty-Guided Adversarial Enhanced Domain Network (UAAD)

[0149] (1) The network optimization problem is expressed as:

[0150]

[0151]

[0152] in, Indicates the optimized parameters of the generated module; represents the optimized parameters of the out-of-domain auxiliary module; Represents the optimized parameters of the diagnostic task module; L cons represents the loss function of the constraint condition; θ G According to L cons Optimize the parameters of WAE; L rela represents the loss function of the relaxation process; θ D According to L rela Optimize the parameters of multi-scale random convolutional networks; L task represents the diagnosis task loss; θ T According to L task Optimize the parameters of the diagnostic task module; represents the optimized parameters of the domain discriminator; represents the optimized parameters of the domain feature extractor; represents the parameters of the optimized domain discriminator; θ F represents the parameters of the optimized domain feature extractor; E(·) represents the calculation of mathematical expectation; θ C represents the optimized classifier;

[0153] (2) The Adam optimizer is used to solve the network optimization problem and update the network parameters in each training round. Specifically, it is expressed as:

[0154]

[0155]

[0156]

[0157] Among them, μ represents the learning rate; represents the generation module parameters in the e-th iteration update process; represents the parameters of the generated module in the e+1th iteration update process after e iteration updates; represents the out-of-domain auxiliary module parameters in the e-th iteration update process; represents the out-of-domain auxiliary module parameters in the e+1th iterative update process after e iterative updates; Represents the diagnostic task module parameters in the e-th iteration update process; represents the diagnostic task module parameters in the e+1th iteration update process after e iteration updates; L CDB represents the loss function that optimizes the class decision boundary; represents the class decision boundary corresponding to the kth fault category in the eth iterative update process; represents the class decision boundary corresponding to the kth fault category in the e+1th iterative update process after e iterations; λ represents the balance L cons Hyperparameters; β represents the balance L rela The hyperparameter υ represents the weight parameter for measuring the decision boundary during the training process of the uncertainty domain adversarial network.

[0158] 4. During the testing phase

[0159] Get the target domain D under unknown conditions t The sensor operation data in the training data may contain known fault categories and unknown fault categories. The sensor operation data is input into the trained sensor fault diagnosis model. The sensor fault diagnosis model assigns the known classes to the boundary according to the learned class decision boundary and rejects the unknown classes outside the boundary. It is expressed as:

[0160]

[0161] Among them, y t N represents the unknown fault sample label; s represents the number of known samples; k+1 represents the unknown fault category; represents the sample characteristics under unknown fault category; d k+1 represents the similarity estimation function based on unknown fault samples;

[0162] In order to verify the effectiveness of the UAAD method proposed in this invention, the data collected by the temperature sensor during the operation of the flash furnace workshop was used for analysis and verification. Figure 3 shown.

[0163] In the experiment conducted in the embodiment of the present invention, sensor samples were collected under 5 different working conditions, and each working condition was operated at a different position in the flash furnace system. The specific working condition position distribution is as follows: Figure 3 The sample contains 2000 sensor health status samples, 1600 of which are randomly selected for training, and the remaining samples are used for preparation of testing. Each sample contains 256 data points. The experiment also set up 5 fault diagnosis tasks, each of which contains sensor health status and 4 fault modes. The data performance of fault mode and health status is shown as follows Figure 4 As shown, it includes drift faults (see Figure 4 Figure (1)), complete failure (see Figure 4 (2) Figure), the accuracy decreases (see Figure 4 Figure (3)), bias fault (see Figure 4 In (4), they are represented by 1, 2, 3, and 4 respectively, which will be used to evaluate the generalization and robustness of the method for diagnosis tasks under different working conditions. The details of the diagnosis task setting are shown in Table 1.

[0164] Table 1 Diagnostic task setting information

[0165]

[0166] In addition, several currently mainstream related diagnostic methods were selected for comparison. The specific implementation details are as follows:

[0167] M1: 1DCNN-SVM. Train a 1DCNN model on the source domain and use SVM to perform diagnosis tasks directly in the unknown target domain.

[0168] M2: DT-DDG. DT-DDG is a relatively new DGFD method in recent years. It combines the adversarial training scheme of dynamic weighting strategy and batch spectral penalty regularization term to extract domain-invariant features and discriminant features from multiple different source domains, and generalizes them to unknown working conditions for diagnostic tasks.

[0169] M3: AMINet. AMINet is a relatively advanced SDGFD method that uses mutual information iterative minimum-maximum game to learn generalized features to resist unknown domain transfer.

[0170] D1: w / oRelax. The relaxation process is removed in the method proposed in the present invention to verify the effectiveness of the method under the same distribution.

[0171] D2: w / oUncertain. In the method proposed in this invention, the uncertainty generalization estimate is removed to evaluate the effectiveness and generalization of the enhanced domain.

[0172] A: UAAD (Proposed). It is the uncertainty-guided adversarial enhancement domain network for single-domain generalized fault diagnosis proposed by the present invention.

[0173] The results of the SDGFD method in performing diagnostic tasks on the flash furnace temperature sensor dataset are shown in Table 2. It can be seen that the method proposed in the present invention is generally better than the comparison method, with an average accuracy of 90%. This shows that UAAD can not only correctly classify known fault modes and sensor health states under unknown working conditions, but also classify unknown fault modes. Among them, in the diagnostic task settings of T3 and T5, the source domain only contains 2 types of known fault modes, and the unknown target domain also contains another 2 unknown fault models, and the performance of the proposed method is still better than several other comparison methods. In order to intuitively reflect the diagnostic results of the proposed method, a bar graph is drawn as follows Figure 5 shown.

[0174] Table 2 H-score of diagnostic tasks (%)

[0175]

[0176]

[0177] In order to further analyze the diagnostic performance of different methods more intuitively, the t-distributed stochastic neighbor embedding (t-SNE) algorithm is used to perform two-dimensional visualization of the features learned by the model, as shown in the figure. Figure 6 The figure shows the proposed method and M1 (see Figure 6 (a) in the figure), M2 (see Figure 6 (b)), M3 (see Figure 6 (c) in the diagnostic task T3 (see Figure 6 The visualization results in (d) are shown in Figure 1. Different colors represent different domains, and different shapes of markers represent different sample modes. M2 and M3 can effectively identify the healthy state samples of the sensor, but when the target unknown fault mode and the target known fault mode samples are mixed, the classification will be poor. M1 cannot divide the boundaries of samples of each mode, and cannot distinguish between known and unknown fault modes, so its classification performance is poor. Figure 6 As can be seen from Figure (d), the method provided by the present invention has a good separation effect on target unknown fault mode and target known fault mode samples, and different unknown fault modes can also be well distinguished. This is because the enhanced domain constructed by the method proposed in this paper has good robustness, which enables the model to effectively generalize to unknown working conditions and correctly classify the sample features of unknown fault modes.

[0178] In order to further verify the effectiveness of the enhanced domain, Figure 7The feature distribution diagram of the enhanced domain on the other four diagnostic tasks is given in Figure 7 (a)), T2 (see Figure 7 (b)), T4 (see Figure 7 (c)), T5 (see Figure 7 (d) in the figure). It can be seen that some features are not well clustered, especially in T2, where the failure mode and unknown mode cannot obtain a clear decision boundary. In other diagnosis tasks, facing unknown target domains, the diagnosis task can still be completed well. This is due to the existence of the enhanced domain, which improves the model's vision and can more clearly find the clear decision boundary of each failure mode, making it easier to diagnose.

[0179] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0180] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A sensor fault diagnosis method based on single domain generalization under uncertainty-guided adversarial enhancement domain, characterized in that: The steps include: Acquire operating data of a target sensor; input the operating data into a trained sensor fault diagnosis model, and output a fault category corresponding to the operating data; The training process of the sensor fault diagnosis model includes a forward propagation process and a back propagation process; In the forward propagation process, an adversarial source domain and a fictitious domain are generated based on a single source domain of a sensor; In the back-propagation process, uncertainty estimation is performed on the adversarial source domain and the imaginary domain to generate an enhanced domain; The sensor fault diagnosis model is trained by the enhanced domain to learn the class decision boundary.

2. The sensor fault diagnosis method based on single-domain generalization under uncertainty-guided adversarial enhancement domain according to claim 1 is characterized in that: The sensor fault diagnosis model is constructed by an uncertainty-guided adversarial enhancement domain network; The uncertainty guided adversarial enhancement domain network includes a generation module, an out-of-domain auxiliary module and an uncertainty domain adversarial network; The out-of-domain auxiliary module is constructed by a multi-scale random convolutional network; The uncertainty domain adversarial network includes a domain feature extractor, a domain identifier and a diagnosis task module.

3. The sensor fault diagnosis method based on single-domain generalization under uncertainty-guided adversarial enhancement domain according to claim 2 is characterized in that: The forward propagation process specifically includes: Acquire a single source domain of a sensor; the single source domain includes a source sample and a fault category label corresponding to the source sample; the source sample includes a plurality of historical operation data of the sensor; In the generation module, using Wasserstein distance as a constraint condition, an adversarial sample consistent with the semantic information of the source sample is generated, and an adversarial source domain is constructed according to the adversarial sample; In the out-of-domain auxiliary module, the source samples are projected into the drift domain as a relaxation process to generate pseudo samples; and a fictitious domain is constructed based on the pseudo samples as a learning object of the generation module.

4. The sensor fault diagnosis method based on single-domain generalization under uncertainty-guided adversarial enhancement domain according to claim 3 is characterized in that: The loss function corresponding to the constraint condition is expressed as: Among them, L cons represents the loss function of the constraint condition; θ G According to L cons Optimize the parameters of WAE; WAE stands for Wasserstein distance-based autoencoder; D s represents a single source domain; D s+ represents the adversarial source domain; d(·) represents the similarity estimation function; ρ represents the maximum domain difference; E(·) represents the calculation of mathematical expectation; W d represents Wasserstein distance; G represents generation module; En represents encoder; x i represents the i-th historical running data of the input; De represents the decoder; z i represents the i-th latent space feature; p(z) represents the latent space feature z i The prior distribution of .

5. The sensor fault diagnosis method based on single-domain generalization under uncertainty-guided adversarial enhancement domain according to claim 4 is characterized in that: The loss function corresponding to the relaxation process is expressed as: Among them, L rela represents the loss function of the relaxation process; θ D According to L rela Optimize the parameters of multi-scale random convolutional networks; D s represents a single source domain; D f represents the fictitious domain; d(·) represents the similarity estimation function; ρ represents the maximum domain difference; D(·) represents the calculation of variance; x s represents the source sample; represents the metric distribution function; p(z) represents the latent space feature z i The prior distribution of .

6. The sensor fault diagnosis method based on single-domain generalization under uncertainty-guided adversarial enhancement domain according to claim 5 is characterized in that: The back propagation process specifically includes: Extracting features from the adversarial source domain and the imaginary domain by the domain feature extractor to obtain a high-level distribution representation of the domain; Inputting the high-level distribution representation into the domain identifier and outputting a confidence logarithm; generating a confidence estimate value by applying the confidence logarithm to an S-curve; Input the high-level distribution representation into the diagnostic task module and output the predicted logarithm; apply the predicted logarithm to the Softmax function to obtain the class prediction probability; Based on the confidence estimate and the class prediction probability, uncertainty estimation is performed on the adversarial source domain and the imaginary domain to generate enhanced samples; and an enhanced domain is constructed according to the enhanced samples; The diagnostic task module is trained through the enhanced domain so that the diagnostic task module learns the class decision boundary.

7. The sensor fault diagnosis method based on single-domain generalization under uncertainty-guided adversarial enhancement domain according to claim 6 is characterized in that: The step of performing uncertainty estimation on the adversarial source domain and the imaginary domain based on the confidence estimate value and the class prediction probability to generate an enhanced sample; constructing an enhanced domain according to the enhanced sample; Specifically include: Taking the correct fault classification label as a hint, we adjust between the confidence estimate c and the class prediction probability p by interpolation to obtain the optimized class prediction probability; specifically, it is expressed as: p i ′=c·p i +(1-c)y i Among them, p i Represents x i The corresponding class prediction probability; x i represents the i-th historical running data input; p i ′ indicates that p i The optimized class prediction probability; c represents the confidence estimate; y i Indicates the value of the i-th element of the fault category label in the one-hot encoding, i.e. x i The corresponding fault category label; Based on the optimized class prediction probability, the diagnostic task loss of the diagnostic task module is calculated; specifically, it is expressed as: Among them, L task represents the diagnostic task loss; M represents the total amount of input historical operation data; Based on the loss function of the constraint condition, the loss function of the relaxation process, and the diagnostic task loss, the total task loss function of the uncertainty-guided adversarial enhancement domain network is obtained; specifically expressed as: L UAAD =L task (i T ;x s )-λL cons (i G ;z)+βL rela (i D ;x s ) Among them, L UAAD represents the total task loss function of the uncertainty-guided adversarial enhancement domain network; θ T represents the parameters of the optimized diagnosis task module; z represents the latent space feature; λ represents the balance L cons Hyperparameters; β represents the balance L rela Hyperparameters of Based on the total task loss function L UAAD , forming enhanced samples and constructing enhanced domains; specifically expressed as: Among them, θ G,D,T Represents θ G ,θ D and θ T ; represents the source sample after t enhancement iterations; z t represents the latent space features after t enhancement iterations; represents the adversarial sample after t enhancement iterations; Indicates that Input into the uncertainty domain adversarial network for enhanced iteration; η represents the learning rate of gradient ascent; represents the enhanced sample after t enhancement iterations; Indicates based on Constructed enhancement domain.

8. The sensor fault diagnosis method based on single domain generalization under uncertainty-guided adversarial enhancement domain according to claim 6 is characterized in that: The step of training the diagnostic task module by using the enhanced domain so that the diagnostic task module learns the class decision boundary specifically includes: The diagnostic task module is enabled to learn sample features of each fault category in the single source domain, and the average feature of each fault category is used as the feature center; The samples in the single source domain and the enhanced domain are input into the diagnostic task module for classification, and the distance between each correctly classified sample feature and the average feature under the same fault category is calculated; specifically expressed as: in, represents the distance between the correctly classified sample features and the average features under the kth fault category; K represents a total of K fault categories; L CE represents the cross entropy function; represents the sample features of the kth fault category in the adversarial source domain; express Fault category label of J k represents the feature center under the kth fault category; C represents the classifier; In the diagnostic task module, the distribution difference of the domain sample space is measured by the conditional maximum average deviation to optimize the distance between each correctly classified sample feature and the average feature under the same fault category; specifically, it is expressed as: Among them, L cmmd represents the loss function of the optimized distance; i represents the i-th historical running data; φ represents the kernel density function; represents the sample features of the kth fault category in the enhanced domain; A corresponding class decision boundary is constructed for each known fault category so that the diagnostic task module can identify known faults and unknown faults; specifically, it is expressed as: in, represents the class decision boundary corresponding to the k-th fault category.

9. The sensor fault diagnosis method based on single domain generalization under uncertainty-guided adversarial enhancement domain according to claim 8 is characterized in that: Also includes: The class decision boundary of each fault category is optimized, which is specifically expressed as: Among them, L CDB represents the loss function that optimizes the class decision boundary; θ k represents the variables that control the effects of the two constraints; represents the decision boundary of the kth fault class.

10. The sensor fault diagnosis method based on single domain generalization under uncertainty-guided adversarial enhancement domain according to claim 2 is characterized in that: It also includes optimizing the uncertainty guided confrontation enhancement domain network; specifically: The network optimization problem is expressed as: in, Indicates the optimized parameters of the generated module; represents the optimized parameters of the out-of-domain auxiliary module; Represents the optimized parameters of the diagnostic task module; L cons represents the loss function of the constraint condition; θ G According to L cons Optimize the parameters of WAE; L rela represents the loss function of the relaxation process; θ D According to L rela Optimize the parameters of multi-scale random convolutional networks; L task represents the diagnosis task loss; θ T According to L task Optimize the parameters of the diagnostic task module; represents the optimized parameters of the domain discriminator; represents the optimized parameters of the domain feature extractor; represents the parameters of the optimized domain discriminator; θ F represents the parameters of the optimized domain feature extractor; E(·) represents the calculation of mathematical expectation; θ C represents the optimized classifier; The Adam optimizer is used to solve the network optimization problem and update the network parameters in each training round. The specific expression is: Among them, μ represents the learning rate; represents the generation module parameters in the e-th iteration update process; represents the parameters of the generated module in the e+1th iteration update process after e iteration updates; represents the out-of-domain auxiliary module parameters in the e-th iteration update process; represents the out-of-domain auxiliary module parameters in the e+1th iterative update process after e iterative updates; Represents the diagnostic task module parameters in the e-th iteration update process; represents the diagnostic task module parameters in the e+1th iteration update process after e iteration updates; L CDB represents the loss function that optimizes the class decision boundary; represents the class decision boundary corresponding to the kth fault category in the eth iterative update process; represents the class decision boundary corresponding to the kth fault category in the e+1th iterative update process after e iterations; λ represents the balance L cons Hyperparameters; β represents the balance L rela The hyperparameter υ represents the weight parameter for measuring the decision boundary during the training process of the uncertainty domain adversarial network.

Images