An industrial control valve fault diagnosis method based on cloud variational autoencoder and integrated geometry construction network

Abstract

The application discloses an industrial control valve fault diagnosis method based on a cloud variation autoencoder and an integrated geometric structure network, and specifically comprises the following steps: collecting control valve small sample fault data, performing uncertainty representation through a two-dimensional cloud model, constructing a cloud characteristic space characterized by expectation, entropy and hyper entropy, and realizing data enhancement based on a general normal distribution; designing a cloud variation autoencoder, optimizing a generation process through a multi-objective loss function, and ensuring that the enhanced data simultaneously satisfy the cloud model general normal distribution characteristics and the distribution difference minimization constraint; utilizing a conic geometric volume to mine the inter-class overlapping degree of the two-dimensional cloud model, providing auxiliary information for fault diagnosis; adopting a geometric structure network to establish a lightweight SAMME integrated learning fault diagnosis framework, fusing the overlapping degree information to construct a strong classifier, and finally outputting a diagnosis result according to an output decision function. The application effectively solves the inter-class feature overlapping problem of small sample fault data, and significantly improves the performance of the control valve fault diagnosis model.

Description

Technical Field

[0001] This invention belongs to the field of control valve fault diagnosis technology, specifically relating to an industrial control valve fault diagnosis method based on cloud variational autoencoder and integrated geometric network. Background Technology

[0002] Control valves, as key components of industrial control systems, typically consist of three parts: an actuator, a valve body, and a smart positioner. Their primary function is to regulate the flow rate, pressure, or level of the medium based on control signals to ensure stable operation of industrial processes. Because control valves often operate under extreme conditions such as high temperature and high pressure, and are frequently exposed to corrosive, toxic, or radioactive media, they are highly susceptible to failure. These failures can not only reduce production efficiency but also lead to production line shutdowns. Therefore, ensuring the stable operation of control valves is crucial for the safety and efficiency of industrial systems.

[0003] To achieve the digital and intelligent transformation of industrial processes, many manufacturers have significantly improved the equipment's status acquisition capabilities by integrating multiple types of sensors into control valves. Based on the multidimensional data collected by sensors and combined with intelligent algorithms, data-driven fault diagnosis models can be constructed. Data-driven fault diagnosis methods do not rely on complex mechanistic models and signal patterns; they rely solely on input and output data to detect potential abnormal faults, greatly promoting the rapid development of the industrial fault diagnosis field.

[0004] However, building a high-precision data-driven fault diagnosis model requires two essential prerequisites: sufficient fault data and data covering different fault types. In industrial processes, equipment often operates under normal conditions. Coupled with the complexity of industrial systems, insufficient equipment informatization, and the instability of the detection environment, these unfavorable factors make it difficult to collect enough fault data, leading to the common small sample size problem in the practical application of data-driven fault diagnosis methods. Furthermore, due to significant overlap in features between fault data categories and blurred feature boundaries between different fault modes, traditional methods struggle to accurately distinguish fault types, further reducing diagnostic accuracy. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes an industrial control valve fault diagnosis method based on cloud variational autoencoders and integrated geometric construction networks to solve the problem of inter-class feature overlap under small sample fault data. This method can achieve high-precision fault diagnosis under limited data conditions, providing reliable technical support for intelligent operation and maintenance of industrial processes.

[0006] To achieve the above objectives, this invention discloses a fault diagnosis method for industrial control valves based on cloud variational autoencoders and integrated geometric construction networks. The method includes the following steps: S1, for small sample fault data of control valve stem displacement and medium flow rate, extracting cloud feature parameters (expected values) using a two-dimensional inverse cloud generator. ,entropy hyperentropy S1) To complete the quantitative characterization of fault uncertainty; S2) Based on the cloud feature parameters, a two-dimensional forward cloud generator is used to map them into the cloud feature space, generating a two-dimensional cloud droplet data set that conforms to a generalized normal distribution, forming an enhanced fault dataset, and realizing the effective expansion of small sample data; S3) To construct a cloud variational autoencoder that integrates cloud feature parameters, the mean and variance of the latent variable distribution are introduced into the cloud feature parameter constraints in the encoder, and the optimized enhanced data is generated through the decoder; S4) To design the loss function of the cloud variational autoencoder, including reconstruction loss, cloud distribution KL divergence loss and maximum mean difference loss optimization generation process, to ensure that the enhanced data simultaneously satisfies the generalized normal distribution characteristics of the cloud model and the constraint of minimizing distribution difference; S5) To perform multi-dimensional visualization analysis on the enhanced fault data based on the two-dimensional cloud model, and to analyze the inter-fault class re-reconstruction... The overlapping region is modeled as a three-dimensional conical surface. The parameters of each cone are obtained through the 3En criterion to construct a quantifiable and analytical geometric model of the overlap. S6: The overlap between fault classes is mined based on the conical geometric volume method, providing more effective decision information for overlapping features that are difficult to distinguish. S7: The SAMME ensemble learning framework is constructed, and the sample weights are dynamically adjusted using an adaptive weight update strategy. It adaptively focuses on samples that are difficult to classify and assigns confidence weights to each base classifier, thereby improving the overall performance of the model. S8: A lightweight geometrically constructed network is used as a weak classifier. Spatial geometric information is embedded in the network structure to effectively reduce modeling costs while maintaining high diagnostic accuracy. S9: The overlap is fused, and the weights and output probabilities of each weak classifier are used to establish the final strong classifier. Finally, a weighted voting strategy function outputs reliable and comprehensive diagnostic results.

[0007] Furthermore, step S1 extracts the valve stem displacement cloud feature parameters using a two-dimensional inverse cloud generator. , , ) and characteristic parameters of medium flow cloud ( , , ),in , , In the formula, This is a small sample of fault data, where n represents the number of small samples.

[0008] Further, step S2 input: cloud feature parameters ( , , , , , Given: 2D cloud data, N is the total number of samples; Output: 2D cloud data. its membership degree , of which Valve stem displacement data For media flow data, N is the total number of two-dimensional cloud data samples; the specific steps are as follows: S21, generate a sample with an expected value of ( , ), variance is ( , Two-dimensional Gaussian random numbers () , S22, produces an expectation of ( , ), with variance ( , Two-dimensional Gaussian random numbers () , S23, Calculate membership degree ,in S24. Repeat steps S21 to S23 until N two-dimensional cloud data are generated.

[0009] Furthermore, step S3 constructs a cloud variational autoencoder that integrates cloud feature parameters, introducing cloud feature parameters as the mean of the latent variable distribution in the encoder. and variance Output constraints are introduced in the variance calculation to extend the standard normal distribution to a generalized normal distribution, and the optimized and enhanced data is generated through the decoder.

[0010] Furthermore, the loss function of the cloud variational autoencoder is designed, including the reconstruction loss, cloud distribution KL divergence loss, and maximum mean difference loss optimization generation process, so that the optimized augmented data satisfies the generalized normal distribution characteristics of the cloud model and the constraint of minimizing distribution difference; step S4 includes the following steps: S41, calculate the reconstruction loss. In the formula, N is the total number of samples. For two-dimensional cloud data, For the generated data; S42, calculate the KL divergence loss of the cloud distribution. In the formula and From step S3, we can obtain that , ,in, At the same time, , As an example of the uncertainty and fuzziness of the pannormal distribution; S43, calculate the maximum mean difference loss. In the formula, n and N represent the total number of small-sample fault data and optimized augmented data samples, respectively. This represents the Gaussian kernel function, where x represents a small sample of fault data. To optimize and augment the data; S44, construct the loss function of the final cloud variational autoencoder. In the formula, and This is the loss coefficient.

[0011] Furthermore, the optimized and enhanced data is mapped into a two-dimensional cloud feature space, and its overlapping region in three-dimensional space is modeled as a conical surface. The radius of the cone's base is calculated using the 3En criterion of the cloud model. It is assumed that the two types of faults are mapped to the two-dimensional cloud feature space as follows: and Step S5 includes the following steps: S51, calculate , and the inner boundary of the overlapping area , , S52, Calculation , and the outer boundary of the overlapping area , , S53, Calculate the radius of the base of each cone. , , S54, Two-dimensional cloud model and The height of the cone is , Calculate the coordinates of the center of the base of the overlapping cone. Substituting this into step S23, the membership degree is obtained, which is the height of the overlapping cone. .

[0012] Furthermore, step S6 includes the following steps: S61, quantifying the overlap between two-dimensional cloud models according to the conical geometric volume method, and calculating respectively , Overlapping areas Approximate cone volume , , S62, using the intersection-union ratio to map the overlap to the range of 0~1, yields... , Final overlap .

[0013] Furthermore, step S7, constructing the SAMME ensemble learning framework, includes the following steps: S71, initializing sample weights: S72, Training a weak classifier with sample weights. S73, Calculate the weighted error rate In the formula, This is an indicator function; it returns 1 if the classification is incorrect, and 0 otherwise. S74 calculates the weights of the weak classifier. In the formula, K is the number of fault categories; S75, update the sample weights. In the formula, This is the normalization factor.

[0014] Furthermore, step S8 includes the following steps: S81, generating within a random interval Set the parameters of candidate nodes and construct a candidate node pool;

[0015] S82 establishes a hidden layer parameter generation mechanism with angle as the control strategy in the network structure. In the formula, Residual With newly added node parameters The angle between them, dynamic parameters , S83, Calculate the hidden layer output weights In the formula, , This represents the output matrix of the Lth node. Indicates generalized inverse, S84 is the output matrix generated by one-hot encoding of the sample labels; S84 calculates the current network residual. S85: Determine whether the network residuals and the number of hidden layer nodes meet the modeling stopping condition; S86: Calculate the final network model output. .

[0016] Furthermore, step S9 integrates the overlap and the weights of each weak classifier to establish a weighted voting strategy function. The output provides reliable and comprehensive diagnostic results. In the formula, for each category k, the sum of the weights of all weak classifiers supporting k categories is calculated, and the category with the largest sum of weights is selected as the final prediction result. This is an indicator function; it returns 1 if the classification is incorrect, and 0 otherwise. T is the total number of weak classifiers. These are the weighting coefficients.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] First, the control valve fault diagnosis method based on cloud variational autoencoder and ensemble learning of the present invention uses a two-dimensional cloud model to characterize the uncertainty of fault data and augment the data, while designing a cloud variational autoencoder. During the encoding process, cloud feature parameters are introduced and multi-objective loss is used to make the augmented data satisfy the generalized normal distribution characteristics of the cloud model and the constraint of minimizing the distribution difference.

[0019] Second, the control valve fault diagnosis method based on cloud variational autoencoder and ensemble learning of the present invention maps fault data of each category to a two-dimensional cloud feature space and models it as a three-dimensional conical surface. It proposes a conical geometric volume method to mine the overlap between different types of faults, providing more effective decision information for fault differentiation.

[0020] Third, the control valve fault diagnosis method based on cloud variational autoencoder and ensemble learning of the present invention constructs a lightweight SAMME ensemble learning fault diagnosis framework, and uses a lightweight geometric construction network as a weak classifier, which effectively reduces modeling costs and maintains high diagnostic accuracy. Attached Figure Description

[0021] Figure 1 This is a flowchart of the control valve fault diagnosis method based on cloud variational autoencoder and ensemble learning according to the present invention.

[0022] Figure 2 This is a schematic diagram of a two-dimensional cloud model generator.

[0023] Figure 3 A schematic diagram showing the mapping of two types of fault data to the overlapping area of ​​the cloud feature space.

[0024] Figure 4 A heatmap showing the overlap between various faults.

[0025] Figure 5 This is a test confusion matrix diagram based on the diagnostic results of the geometrically constructed network.

[0026] Figure 6 This is a test confusion matrix diagram of the diagnostic results of the method proposed in this invention. Detailed Implementation

[0027] The following embodiments are provided to enable those skilled in the art to more fully understand the present invention, but do not limit the invention in any way.

[0028] like Figure 1 As shown, this invention proposes a fault diagnosis method for industrial control valves based on cloud variational autoencoders and integrated geometric construction networks, comprising the following steps:

[0029] S1: The small sample fault data of control valve stem displacement and medium flow obtained are processed through methods such as... Figure 2 The two-dimensional inverse cloud generator shown extracts its cloud feature parameters (expected). ,entropy hyperentropy ), to complete the quantitative characterization of fault uncertainty; step S1 is as follows:

[0030] Input: Small sample fault data ;

[0031] Output: Expected ,entropy hyperentropy ;

[0032] S11: Calculate the expectation ;

[0033] S12: Calculate entropy ;

[0034] S13: Calculate hyperentropy ;

[0035] Based on the above formula, the characteristic parameters of the valve stem displacement cloud are obtained respectively. , , ) and characteristic parameters of medium flow cloud ( , , ).

[0036] S2: Based on the cloud feature parameters, a two-dimensional forward cloud generator is used to map them into the cloud feature space, generating a two-dimensional cloud droplet data set that conforms to a generalized normal distribution, forming two-dimensional cloud data, and realizing the effective expansion of small sample data; step S2 is as follows:

[0037] Input: Two-dimensional cloud data its membership degree ,in For valve stem displacement data, For media flow data, N is the number of two-dimensional cloud data samples.

[0038] Output: Valve stem displacement enhancement data Media flow enhancement data its membership degree , .

[0039] S21: Generate an expected value of ( , ), variance is ( , Two-dimensional Gaussian random numbers () , );

[0040] S22: Generate an expectation of ( , ), with variance ( , Two-dimensional Gaussian random numbers () , );

[0041] S23: Calculate membership degree ,in ;

[0042] S24: Repeat steps S21 to S23 until N two-dimensional cloud data are generated;

[0043] Based on the above steps, the final two-dimensional cloud data is obtained. This effectively characterizes the inherent randomness of the original fault data.

[0044] S3: Construct a cloud variational autoencoder that integrates cloud feature parameters. In the encoder, the mean and variance of the latent variable distribution are introduced into the cloud feature parameter constraints, and the decoder generates optimized and enhanced data. Step S3 is detailed below:

[0045] A cloud variational autoencoder is designed by fusing cloud feature parameters, and cloud feature parameters are introduced into the encoder as the mean of the latent variable distribution. and variance Output constraints, namely:

[0046]

[0047] In variance calculation, hyperentropy He is introduced to extend the standard normal distribution to a generalized normal distribution, and optimized and enhanced data is generated through a decoder;

[0048] S4: Design the loss function of the cloud variational autoencoder, including the reconstruction loss, cloud distribution KL divergence loss, and maximum mean difference loss optimization generation process, to ensure that the augmented data simultaneously satisfies the generalized normal distribution characteristics of the cloud model and the constraint of minimizing distribution difference; Step S4 is as follows:

[0049] S41: Calculate the reconstruction loss This reflects the difference between the generated data and the input, with the goal of making the generated result as similar as possible to the input two-dimensional cloud data. Its calculation formula is as follows:

[0050]

[0051] In the formula, N is the total number of optimized and augmented data samples. For two-dimensional cloud data, To enhance the data after optimization;

[0052] S42: Calculate the KL divergence loss of cloud distribution. The goal is to make the latent variables generated by the encoder conform as closely as possible to the generalized normal distribution of the cloud model. The calculation formula is as follows:

[0053]

[0054] In the formula and From step C, we can obtain that , ,in, At the same time, , As a result of the uncertainty and ambiguity of the generalized normal distribution;

[0055] S43: Calculate the distribution distance loss The goal is to minimize the distribution difference between the generated data and the original fault data, and its calculation formula is as follows:

[0056]

[0057] In the formula, n and N represent the total number of small-sample fault data and augmented data samples, respectively. This represents the Gaussian kernel function, where x represents a small sample of fault data. To enhance the data;

[0058] S44: Construct the loss function for the final cloud variational autoencoder:

[0059]

[0060] In the formula, and The loss coefficient, , and These represent the reconstruction loss, cloud distribution KL divergence loss, and distribution difference loss, respectively.

[0061] S5: As Figure 3 As shown, the enhanced fault data is visualized and analyzed based on a two-dimensional cloud model. The overlapping areas between fault classes are modeled as three-dimensional conical surfaces. The parameters of each cone are obtained through the 3En criterion, and a quantifiable and analytical geometric model of the overlap is constructed. Step S5 is as follows:

[0062] The reconstructed fault data of different categories are mapped to a two-dimensional cloud feature space, which approximates a conical surface in three-dimensional space. The radii of the cone's base are calculated using the 3En criterion of the cloud model. Assuming the two types of faults are mapped to the two-dimensional cloud feature space as follows... and Its degree of overlap The calculation steps are as follows:

[0063] S51: Calculation , And the inner boundary of the overlapping region:

[0064]

[0065] S52: Calculation , And the outer boundary of the overlapping region:

[0066]

[0067] S53: Calculate the radius of the base of each cone , , :

[0068]

[0069] S54: Two-dimensional cloud model and The height of the cone is , Calculate the coordinates of the center of the base of the overlapping cone. Substituting this into step S23, the membership degree is obtained, which is the height of the overlapping cone. .

[0070] S6: Based on the conical geometric volume method, the overlap between fault categories mapped to the two-dimensional cloud feature space is mined to provide more effective decision information for overlapping features that are difficult to distinguish; Step S6 is as follows:

[0071] S61: Quantify the overlap between two-dimensional cloud models using the conical geometric volume method, and calculate respectively. , Overlapping areas Approximate volume of the cone:

[0072]

[0073] S62: The overlap is mapped to the range of 0 to 1 using the crossover-union ratio, resulting in... , Final overlap:

[0074]

[0075] S7: Construct the SAMME ensemble learning framework, employing an adaptive weight update strategy to dynamically adjust sample weights, adaptively focusing on samples that are difficult to classify, and allocating confidence weights to each base classifier, thereby improving the overall performance of the model; Step S7 is detailed below:

[0076] S71: Initialize sample weights: ;

[0077] S72: Iteratively train the weak classifier, which includes the following sub-steps:

[0078] S73: Training a weak classifier with sample weights. , where t is the number of weak classifiers;

[0079] S74: Calculate the weighted error rate:

[0080]

[0081] In the formula, It is an indicator function; it returns 1 if the classification is incorrect, and 0 otherwise.

[0082] S75: Calculate the weights of the weak classifier:

[0083]

[0084] In the formula, K represents the number of fault categories;

[0085] S76: Update sample weights:

[0086]

[0087] In the formula, This is the normalization factor.

[0088] S8: A lightweight geometrically constructed network is used as a weak classifier, embedding spatial geometric information into the network structure to effectively reduce modeling costs while maintaining high diagnostic accuracy; step S8 is as follows:

[0089] S81: Generate within a random interval Set the parameters of candidate nodes and construct a candidate node pool;

[0090] S82: A hidden layer parameter generation mechanism with angle as the control strategy was established in the network structure.

[0091]

[0092] In the formula, Residual With newly added node parameters The angle between them, dynamic parameters , .

[0093] S83: Calculate the hidden layer output weights:

[0094]

[0095] In the formula, , This represents the output matrix of the Lth node. Indicates generalized inverse, It is the output matrix generated by one-hot encoding of the sample labels;

[0096] S84: Calculate the current network residual Determine whether the network residuals and the number of hidden layer nodes meet the modeling stopping condition;

[0097] S85: Calculate the final network model output. .

[0098] S9: The final strong classifier is established by fusing the overlap coefficient matrix, the weights of each weak classifier, and the data probabilities. A reliable and comprehensive diagnostic result is then output using a weighted voting strategy function. Step S9 is detailed below.

[0099] A weighted voting strategy function is established to output reliable and comprehensive diagnostic results. The weighted voting strategy function is as follows:

[0100]

[0101] In the formula, for each category k, the sum of the weights of all weak classifiers supporting k categories is calculated, and the category with the largest sum of weights is selected as the final prediction result. This is an indicator function; it returns 1 if the classification is incorrect, and 0 otherwise. T is the total number of weak classifiers. The weighting coefficients are used. To illustrate the performance of this invention in diagnosing inter-class feature overlap faults with a small sample size, an example of an industrial control valve is used. Experiments are conducted using one class of fault-free valves and seven classes of control valve faults. For example... Figure 4 The diagram illustrates the overlap between fault classes. It shows that the proposed method can accurately calculate the overlap between classes; faults f3 and f4, and f6 and f7 all exhibit class overlap. Furthermore, this invention compares the prediction accuracy of fault class prediction using only geometrically constructed networks with that of the proposed control valve fault diagnosis method based on cloud variational autoencoders and ensemble learning. Figure 5 and Figure 6 As can be seen, the prediction accuracy using only geometrically constructed networks is only 0.8375, and it is difficult to distinguish overlapping faults from the confusion matrix. In contrast, the method proposed in this invention has an accuracy as high as 0.9458, and it can also be observed from the confusion matrix that it can effectively distinguish overlapping faults, thereby improving the accuracy of fault diagnosis.

[0102] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for industrial control valves based on cloud variational autoencoders and integrated geometric construction networks, characterized in that, The method includes the following steps: Step A: Collect small sample fault data of valve stem displacement and medium flow rate of control valve, and extract its cloud feature parameters through a two-dimensional reverse cloud generator; Step B: Based on the cloud feature parameters obtained in Step A, a two-dimensional forward cloud generator is used to map them into the cloud feature space to generate a two-dimensional cloud droplet data set that conforms to a generalized normal distribution, thus forming enhanced two-dimensional cloud data. Step C: Construct a cloud variational autoencoder that integrates cloud feature parameters. Take the two-dimensional cloud data obtained in step B as input. In the encoder, the mean and variance of the latent variable distribution are introduced into the cloud feature parameter constraints, and the decoder generates optimized and enhanced data. Step D: Design the loss function of the cloud variational autoencoder, including reconstruction loss, cloud distribution KL divergence loss and maximum mean difference loss optimization generation process, so that the optimized augmented data obtained in step C simultaneously satisfies the generalized normal distribution characteristics of the cloud model and the constraint of minimizing distribution difference. Step E: Based on the two-dimensional cloud model, perform multi-dimensional visualization analysis on the optimized and enhanced data obtained in step D, model the overlapping area between fault classes as a three-dimensional conical surface, obtain the parameters of each cone through the 3En criterion, and construct a quantifiable and analytical geometric model of overlap. Step F: Based on the overlap geometric model constructed in Step E, the conical geometric volume method is used to mine the overlap between fault classes, providing decision information for overlapping features that are difficult to distinguish. Step G: Construct the SAMME ensemble learning framework, using the optimized and augmented data obtained in step C as the model input, dynamically adjusting the sample weights using an adaptive weight update strategy, focusing on samples that are difficult to classify, and configuring confidence weights for each weak classifier. In step H, during the construction of the integrated learning framework in step G, a lightweight geometric construction network is used as a weak classifier, and spatial geometric information is embedded in the network structure. Step I: The overlap mined in Step F, the weights of each weak classifier assigned in Step G, and the output probabilities of each weak classifier in Step H are combined to build the final strong classifier. Finally, the diagnostic results are output by the weighted voting strategy function.

2. The industrial control valve fault diagnosis method based on cloud variational autoencoder and integrated geometry construction network according to claim 1, characterized in that, The cloud feature parameters to be extracted in step A are expectation , entropy , hyper-entropy , specifically extracted by a two-dimensional reverse cloud generator, that is: ， In the formula, This is for a small sample of fault data. , where n represents the total number of small samples; according to the above formula, the characteristic parameters of valve stem displacement cloud can be obtained respectively ( , , ) and characteristic parameters of medium flow cloud ( , , ).

3. The method according to claim 2, wherein, Step B involves inputting cloud feature parameters ( , , , , , Output two-dimensional cloud data its membership degree ,in For valve stem displacement data, For media flow data, N is the number of two-dimensional cloud data samples; repeat the input and output steps until the number of samples is expanded to N, as detailed below: Step B1: Generate an expected value of ( , ), variance is ( , Two-dimensional Gaussian random numbers () , ); Step B2: Generate an expected value of ( , ), with variance ( , Two-dimensional Gaussian random numbers () , ); Step B3: Calculate the membership , where ; Step B4: Repeat steps B1 to B3 until N two-dimensional cloud data are generated.

4. The industrial control valve fault diagnosis method based on cloud variational autoencoder and integrated geometric network according to claim 3, characterized in that, Step C is as follows: A cloud variational autoencoder with fusion cloud feature parameters is constructed, and cloud feature parameters are introduced into the encoder as the mean of the hidden variable distribution and variance Output constraints, namely: ， In variance calculation, hyperentropy He is introduced to extend the standard normal distribution to a generalized normal distribution, and optimized and enhanced data is generated through a decoder.

5. The method according to claim 4, wherein, Step D involves designing the loss function for the cloud variational autoencoder, including the following steps: Step D1: Calculate the reconstruction loss This reflects the difference between the generated data and the input, with the goal of making the generated result as similar as possible to the input two-dimensional cloud data. Its calculation formula is as follows: ， where N is the total number of two-dimensional cloud data samples, is the two-dimensional cloud data, is the enhanced data after optimization; Step D2: Compute cloud distribution KL divergence loss The goal is to make the latent variable generated by the encoder as much as possible to conform to the general normal distribution of the cloud model, and its calculation formula is as follows: ， From step C With , , Wherein, While , as uncertainty and ambiguity of the generalized normal distribution; Step D3: Calculate the maximum mean difference loss The goal is to minimize the distribution difference between the generated data and the original fault data, and its calculation formula is as follows: ， where n and N represent the total number of small sample fault data and optimized enhanced data samples, represents a Gaussian kernel function; x is small sample fault data, is the optimized enhanced data; Step D4: Construct the loss function for the final cloud variational autoencoder: ， wherein with is a loss coefficient, , with are reconstruction loss, cloud distribution KL divergence loss and maximum mean discrepancy loss, respectively.

6. The method according to claim 5, wherein, Step E is as follows: Based on steps A and B, the reconstructed fault data of different categories are mapped into a two-dimensional cloud feature space. This space is then approximated as a conical surface in three-dimensional space. The radii of the cone's base are calculated using the 3En criterion of the cloud model. Assuming the two types of faults are mapped to the two-dimensional cloud feature space as follows... and Their overlapping areas The calculation steps are as follows: Step E1: Calculate , and inner boundaries of the overlap region: ， Step E2: Calculate , and the outer boundary of the overlap region: ， Step E3: Calculate the radius of each conic base , , : ， Step E4: Two-dimensional cloud model and The height of the cone is , Calculate the coordinates of the center of the base of the overlapping cone. Substitute this into step B3 to obtain the membership degree, which is the height of the overlapping cone. .

7. The cloud-based variational autoencoder and ensemble learning based control valve fault diagnostic method of claim 6, wherein, Step F is as follows: According to the conic geometry volume method, the overlap degree between the two-dimensional cloud models is quantified, and the approximate conic volume of the overlapping region O is calculated respectively as follows: , ​ ， The overlap degree is mapped into 0~1 by using the intersection-over-union ratio to obtain , The final overlap degree between the two images .

8. The method according to claim 7, wherein, Step G includes the following steps: Step G1: Initialize sample weights: ; Step G2: Iteratively train the weak classifier, which includes the following sub-steps: Step G21: Train weak classifiers under sample weights where t is the number of weak classifiers. Step G22: Calculate the weighted error rate: ， In the formula, It is an indicator function; it returns 1 if the classification is incorrect, and 0 otherwise. Step G23: Calculate the weights of the weak classifiers: ， In the formula, K is the number of fault categories; Step G24: Update sample weights: ， In the formula, is a normalization factor.

9. The method according to claim 8, wherein, Step H is as follows: Step H1: generating in a random interval group candidate node parameters, and constructing a candidate node pool; Step H2: A lightweight geometrically constructed network is used as a weak classifier, and a hidden layer parameter generation mechanism with angle as the control strategy is established in the network structure: ， In the formula, Residual With newly added node parameters The angle between them, dynamic parameters , ; Step H3: Calculate the hidden layer output weights: ， In the formula, , represents the output matrix of the Lth node, represents the generalized inverse, is the output matrix generated by one-hot encoding of the sample label; Step H4: Calculate the current network residual , determine whether the network residual and the number of hidden layer nodes satisfy the modeling stop condition; Step H5: Calculate the final network model output. .

10. The method according to claim 9, wherein, Step I is as follows: Based on overlap, the weights and output probabilities of each weak classifier are used to construct the final strong classifier. A weighted voting strategy function then outputs reliable and comprehensive diagnostic results, as follows: ， In the formula, for each category k, the sum of the weights of all weak classifiers supporting k categories is calculated, and the category with the largest sum of weights is selected as the final prediction result. This is an indicator function; it returns 1 if the classification is incorrect, and 0 otherwise. T is the total number of weak classifiers. These are the weighting coefficients.

Images