Zero-sample performance degradation evaluation method for intelligent decision-making of water washing of gas compressor

The method addresses the challenge of setting accurate compressor degradation thresholds by using a domain adversarial network and generative adversarial network to generate domain-invariant features, enhancing the precision of performance assessment and water washing decisions in gas-steam combined cycle power generation units.

CN119089272AActive Publication Date: 2024-12-06ZHEJIANG UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411220179.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-02
Publication Date
2024-12-06
Estimated Expiration
2044-09-02

AI Technical Summary

Technical Problem

Existing data-driven methods for compressor performance degradation assessment in gas-steam combined cycle power generation units are inadequate due to the inability to set accurate degradation thresholds under varying load conditions and the lack of high degradation data, leading to inefficient water washing decisions.

Method used

A zero-sample performance degradation assessment method using a domain adversarial structure network and a degraded difference generative adversarial network to extract domain-invariant features and generate high-level degradation features from low-level data, assisted by expert knowledge and attribute vectors.

Benefits of technology

Improves the accuracy of compressor performance degradation assessment by generating domain-invariant features and overcoming the challenge of lacking high degradation data, enabling more precise water washing decisions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089272A_ABST
    Figure CN119089272A_ABST
Patent Text Reader

Abstract

The invention discloses a zero sample performance degradation evaluation method for intelligent decision making of gas compressor water washing, and the method comprises the steps: designing a degradation difference generation network guided by gas compressor degradation knowledge, carrying out the directional degradation of data features with slight performance degradation under the assistance of priori knowledge, and generating data features with severe performance degradation; a classifier is trained in a supervised manner, and whether washing is carried out or not is judged according to the output degradation level; considering that the evaluation of the real degradation degree of the gas compressor is interfered when the performance of the gas compressor is improved along with the rising of the load of the gas turbine, different load working conditions are regarded as multiple domains, and invariable characteristics under each load working condition are extracted from operation data to relieve the interference caused by the load change; and various knowledge information in the features is reserved through a feature extractor, so that the generation quality of the degradation difference generation network is improved. According to the method, the degradation level can be evaluated under the condition that the serious degradation training data of the gas compressor is missing, so that support is provided for intelligent washing decision-making of the gas generator set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment performance degradation assessment, and in particular to a zero-sample performance degradation assessment method for compressor water washing intelligent decision-making. Background Art

[0002] Gas-steam combined cycle power generation units are widely used around the world for their high efficiency and low pollution characteristics. Gas-steam combined cycle power generation units are designed to convert chemical energy into electrical energy. As a key component of the combined cycle power generation system, the performance of the gas turbine is a key factor in determining the operating status of the entire power plant. As the core component of the gas turbine power generation system, the compressor undertakes the important task of compressing air and gas for combustion. However, since the filtration system at the air inlet cannot completely remove all impurities in the air, the dirt accumulated on the compressor blades during long-term operation will reduce its efficiency, increase operating costs, and sometimes even cause surge, which has a serious impact on the equipment. For this reason, it is crucial to regularly perform offline water washing of the compressor to ensure the safety, reliability and environmental protection of the power generation process.

[0003] At present, many power plants still determine the compressor water washing cycle based on the operating time of the unit. However, this fixed time interval cannot accurately reflect whether the actual performance of the compressor has decreased, which leads to waste of water washing resources due to frequent water washing, or compressor efficiency decreases due to untimely water washing. Therefore, more and more researchers have begun to use data-driven methods to study the degree of compressor performance degradation and assist in water washing decisions. For example, some researchers have proposed a compressor benchmark pressure ratio model, and then obtained the benchmark pressure ratio range and calculated the upper and lower limits of the pressure ratio drop. The larger the pressure ratio drop, the higher the compressor fouling degree is considered. When the pressure ratio drop exceeds the water washing threshold, the gas turbine should be arranged for offline water washing; some researchers selected characteristic indicators reflecting the degree of compressor fouling based on mechanism analysis, and established a benchmark value model of characteristic indicators by comprehensively considering the influence of boundary conditions and operating means. The fouling degree calculated based on this characteristic indicator is combined with a priori threshold to assist in offline water washing decisions; some researchers use the deviation between the actual operating values ​​of four performance parameters and their neural network prediction values ​​under ideal conditions to measure the degree of compressor performance degradation, providing a criterion for offline water washing. However, the existing data-driven methods use the threshold method of performance evaluation indicators. However, the compressor performance is highly correlated with the power generation load. Under the highest power operating conditions, the thermodynamic cycle of the gas turbine is closer to the ideal thermodynamic cycle. At this time, the compressor performance will be higher than the low-power operating conditions under the same other conditions. Therefore, the degradation level threshold of the indicator will also change in the scenario of frequent load changes of the gas turbine. Secondly, the compressor measurement point value changes with the load change controlled by the grid side, so the model needs to consider the corresponding relationship between the measurement point value and the compressor performance under different load conditions. In addition, power plants usually do not allow compressors to operate to a serious degree of degradation, so data with a high degree of degradation is difficult to obtain, and it is impossible to set the corresponding degradation threshold based on the data.

[0004] In view of the difficulty in setting thresholds due to the frequent load changes and high degradation degree data missing of gas turbines, an expert knowledge-assisted zero-shot learning strategy is urgently needed to evaluate the degree of compressor performance degradation. Summary of the invention

[0005] The purpose of the present invention is to provide a zero-sample performance degradation assessment method for intelligent decision-making of compressor water washing in view of the deficiencies of the prior art. The present invention proposes a solution to the problem that the performance degradation grade assessment model cannot be established due to the inability to obtain severely degraded data. The present invention explores the gradual characteristics of compressor performance degradation and improves the accuracy of zero-sample performance degradation assessment in industrial processes.

[0006] The object of the present invention is achieved through the following technical solution: a zero-sample performance degradation assessment method for compressor water washing intelligent decision-making, comprising the following steps:

[0007] (1) The performance degradation process of the gas turbine compressor is divided into multiple degradation levels, and the operating data of all degradation levels under various load conditions of the compressor are collected. The training set is constructed using the operating data of the degradation level that has been seen and the degradation level label, load condition label and attribute vector describing the degradation level corresponding to the operating data. The test set is constructed using the operating data of the degradation level that has not been seen and the operating data of the degradation level that has been seen and the degradation level label, load condition label and attribute vector describing the degradation level corresponding to the operating data.

[0008] (2) constructing a domain adversarial structure network, which includes a feature extractor, a gradient reversal layer, a domain adversarial discriminator, and an attribute predictor; using a training set to train the domain adversarial structure network to obtain a trained feature extractor, a domain adversarial discriminator, and an attribute predictor;

[0009] (3) constructing a degraded difference generative adversarial network, which includes the trained feature extractor and attribute predictor, generator, converter, and generative adversarial discriminator obtained in step (2); training the degraded difference generative adversarial network using the training set to obtain the trained generator, converter, and generative adversarial discriminator;

[0010] (4) Based on the training set and the test set, the trained feature extractor obtained in step (2) is used to extract domain-invariant seen degradation level features, and the trained generator and transformer obtained in step (3) are used with the assistance of the attribute vector to generate domain-invariant unseen degradation level features. The domain-invariant seen degradation level features and the domain-invariant unseen degradation level features and their corresponding degradation level labels are used to perform supervised training on the classifier to obtain a trained classifier for all degradation levels;

[0011] (5) When applying, the running data to be tested in the test set is first input into the trained feature extractor to obtain the corresponding domain-invariant features, and then the domain-invariant features are input into the trained classifier to obtain the corresponding degradation level.

[0012] Furthermore, the observed degradation level and the unobserved degradation level are obtained by the following method:

[0013] A 60%-40% partitioning strategy is adopted to divide the multiple degradation levels into seen degradation levels and unseen degradation levels. That is, after all degradation levels are sorted from small to large according to the degree of performance degradation, the first 60% of the degradation levels are selected as seen degradation levels, and the remaining degradation levels are selected as unseen degradation levels.

[0014] Furthermore, the training set is expressed as in, The operating data used for training indicates the degradation level seen during compressor operation, d x Indicates the data dimension, represents the set of running data used for training with the observed degradation level; y tr Represents x tr The corresponding degradation level label, represents the set of labels of the seen degradation levels, K s Indicates the number of degradation levels seen; Represents x tr The corresponding attribute vector, A represents the set of all attribute vectors of all degradation levels; w s Represents x tr The corresponding load condition labels, W = {1, ..., Q} represents the set of load condition labels, and Q represents the number of load conditions;

[0015] The test set is represented as Among them, x te Indicates the running data to be tested. represents the set of operational data for which no degradation level is observed, The set of operational data used for the test that represents the degradation level seen; y te Represents x te The corresponding degradation level label, represents the set of unseen degradation level labels, K is the total number of all degradation levels; a u Represents x te The corresponding attribute vector; w u Represents x te Corresponding load case label.

[0016] Furthermore, the use of the training set to train the domain adversarial structure network to obtain a trained feature extractor, a domain adversarial discriminator, and an attribute predictor specifically includes:

[0017] The operating data under each load condition in the training set is input into the domain adversarial structure network. The features of the operating data are first extracted by the feature extractor. The features extracted by the feature extractor are then passed through the gradient reversal layer and the domain adversarial discriminator in turn to obtain the probabilities corresponding to each domain label. The domain label corresponding to the maximum probability is selected as the predicted domain label output by the domain adversarial discriminator. The features extracted by the feature extractor are sent to the attribute predictor to obtain the corresponding attribute vector prediction value.

[0018] The cross entropy loss function is used to calculate the domain discrimination loss function according to the predicted domain labels of the running data and the corresponding true domain labels in the training set; the binary cross entropy loss function is used to calculate the first label discrimination loss function according to the predicted value of the attribute vector output by the attribute predictor and the corresponding true value of the attribute vector in the training set; the domain discrimination loss function and the first label discrimination loss function are weighted to obtain the total loss function of the domain adversarial structure network, and the parameters of the feature extractor, domain adversarial discriminator and attribute predictor are adjusted with the minimization of the total loss function of the domain adversarial structure network as the optimization goal to obtain the trained feature extractor, domain adversarial discriminator and attribute predictor; wherein the trained feature extractor can extract domain invariant features with attribute differences.

[0019] Furthermore, the calculation formula of the domain discrimination loss function is:

[0020]

[0021] Among them, L D represents the domain discrimination loss function, f(·) represents the features output by the feature extractor, and D a (·) represents the predicted domain label output by the domain adversarial discriminator, w s represents the true domain label and also the load condition label. P(·) represents the probability that the predicted domain label is the true domain label. It represents the mathematical expectation and the average loss on the training set;

[0022] The calculation formula of the first label discrimination loss function is:

[0023]

[0024] Among them, L A represents the first label discrimination loss function, M represents the total number of attributes in the attribute vector, and A(·) j represents the predicted value of the jth attribute output by the attribute predictor, a s,j Represents data x tr The true value of the jth attribute in the corresponding attribute vector.

[0025] Furthermore, the use of the training set to train the degraded difference generative adversarial network to obtain a trained generator, a converter and a generative adversarial discriminator specifically includes:

[0026] First, randomly sample all samples in the training set and take out two samples Constitute a sample pair, where And assume Then calculate the high-level samples The attribute vector With low-level samples The attribute vector The difference between them is taken as the attribute difference, and then the attribute difference is spliced ​​with random noise and input into the generator to obtain the generated feature difference; at the same time, the high-level samples are Operation data and low-level samples Operation data Input it into the trained feature extractor to obtain the corresponding high-level features and low-level features, and use them as the true high-level features and true low-level features respectively; secondly, add the generated feature difference to the true low-level features and input them into the converter to obtain the generated high-level features; then, add the true high-level features and the generated high-level features to the attribute vector of the high-level sample respectively. After the combination, the high-level features are input into the generative adversarial discriminator to obtain the discrimination result of the high-level features, that is, whether the high-level features input into the generative adversarial discriminator are real or generated; at the same time, the generated high-level features are input into the trained attribute predictor to obtain the corresponding attribute vector prediction value;

[0027] The adversarial loss function is calculated based on the discrimination results corresponding to the real high-level features output by the feature extractor and the generated high-level features output by the converter; the binary cross entropy loss function is used to calculate the attribute vector prediction value corresponding to the high-level features output by the attribute predictor and the attribute vector of the corresponding high-level samples in the training set. The true value is used to calculate the second label discrimination loss function; the generative adversarial loss function and the second label discrimination loss function are weighted to obtain the total loss function of the degraded difference generative adversarial network, and the total loss function of the degraded difference generative adversarial network is minimized as the optimization goal, and the parameters of the generator, converter and generative adversarial discriminator are adjusted to obtain the trained generator, converter and generative adversarial discriminator; wherein the trained generator and converter can use attribute vectors and low-level features to generate high-level features with attribute discrimination.

[0028] Furthermore, the calculation formula of the generated adversarial loss function is:

[0029]

[0030] in, represents the generated adversarial loss function, D b (·) represents the discriminant result of generating high-level features output by the adversarial discriminator, represents the high-level features output by the trained feature extractor, represents the high-level features generated by the converter output, λ represents the penalty coefficient, α~U(0,1) represents the weight coefficient, U(0,1) represents the uniform distribution defined on the interval [0,1], ||·||2 represents the second norm, surface about The gradient of , E represents the mathematical expectation, which represents the average loss on all the sample pairs sampled above;

[0031] The calculation formula of the second label discrimination loss function is:

[0032]

[0033] Among them, L A ′ represents the second label discrimination loss function, Indicates high-level samples Running data in The corresponding attribute vector The true value of the jth attribute in .

[0034] Furthermore, the step (4) specifically includes:

[0035] First, sample the training set data x in the training set tr and its corresponding degradation level label y tr and attribute vector a s , randomly select attribute vector a from the unseen degradation levels u and its corresponding degradation level label y u ; Then use the trained feature extractor to extract the training set data x tr The domain-invariant degradation level features seen are used as training set features; secondly, the unseen degradation level attribute vector a is calculated u The attribute vector a of the training set data s The attribute difference between the two sets is obtained by splicing the attribute difference with random noise and then inputting it into the trained generator to obtain the generated feature difference; the generated feature difference is then added to the training set feature and inputted into the trained converter to obtain the generated unseen degradation level feature; finally, the generated unseen degradation level feature and the training set feature are respectively inputted into the classifier to obtain the corresponding predicted degradation level label;

[0036] Using the cross entropy loss function, according to the unseen degradation level y u The generated features and corresponding predicted degradation level labels, as well as the predicted degradation level labels corresponding to the training set features and the degradation level labels corresponding to the training set data y tr, calculate the supervised training loss function of the classifier; take minimizing the supervised training loss function of the classifier as the optimization goal, adjust the parameters of the classifier, and obtain a trained classifier; wherein the trained classifier can simultaneously classify the seen degradation level features and the unseen degradation level features.

[0037] Furthermore, the calculation formula of the supervised training loss function of the classifier is:

[0038]

[0039] Among them, L C represents the supervised training loss function of the classifier, C(·) represents the predicted degradation level label output by the classifier, G(·) represents the output of the generator, T(·) represents the output of the transformer, and z represents random Gaussian noise.

[0040] The beneficial effects of the present invention are as follows: the present invention proposes a knowledge-data fusion zero-sample performance degradation assessment method for the compressor performance level assessment problem; in addition, since the compressor performance is highly correlated with the power generation load, in order to consider the influence of variable load on the judgment of the compressor performance degradation degree, consistent prior knowledge is designed for each degradation level under each load condition to describe the compressor performance state, and at the same time, different load conditions are regarded as multiple domains, and a domain adversarial structure is introduced to obtain domain invariant features; on the basis of this generative zero-sample learning, the present invention proposes a degradation difference generative adversarial network for the degradation zero-sample problem, introduces the idea of ​​gradual performance degradation, and generates high degradation level features using low degradation level features with the assistance of knowledge, thereby overcoming the problem that high degradation level data is difficult to obtain and cannot be modeled, and improving the quality of generated data; the present invention can mine the gradual characteristics of compressor performance degradation, improve the accuracy of zero-sample performance degradation assessment in industrial processes, and assist compressor water washing decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the zero-sample performance degradation assessment method for compressor water washing intelligent decision-making of the present invention;

[0042] Figure 2 A schematic diagram of the structure of the domain adversarial structure network of the present invention;

[0043] Figure 3 A schematic diagram of the structure of the degenerate difference generation adversarial network of the present invention;

[0044] Figure 4 The present invention is a flow chart of using a feature extractor, a generator and a converter to train a classifier. DETAILED DESCRIPTION

[0045] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0046] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0047] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0048] The present invention is described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0049] This embodiment verifies the effectiveness of the present invention based on measured data of a thermal power plant, but is not limited thereto. Any industrial project with progressive degradation characteristics can use the method of the present invention to perform zero-sample performance degradation level assessment.

[0050] See also Figure 1 The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making of the present invention specifically includes the following steps:

[0051] (1) The performance degradation process of a gas turbine compressor is divided into multiple degradation levels. The operating data of all degradation levels under various load conditions of the compressor are collected. The training set is constructed using the operating data of the seen degradation levels for training and the degradation level labels, load condition labels and attribute vectors describing the degradation levels corresponding to the operating data. The test set is constructed using the operating data of the unseen degradation levels and the operating data of the seen degradation levels for testing and the degradation level labels, load condition labels and attribute vectors describing the degradation levels corresponding to the operating data.

[0052] In this embodiment, the performance degradation process of the gas turbine compressor is divided into K degradation levels. In order to consider the impact of variable load on the compressor performance degradation threshold, expert prior knowledge with constant load is designed for each degradation level to reflect the compressor performance state, wherein the prior knowledge includes M attributes. Experts use four evaluation criteria of "excellent", "good", "medium", and "poor" to measure the attributes of each degradation level. Of course, the selection of the evaluation criteria can be selected according to actual needs, for example, using more detailed 7 evaluation criteria, namely "excellent upper", "lower excellent", "good upper", "lower good", "upper medium", "lower medium", and "poor". A combination of evaluation criteria containing all attributes is called an attribute vector. Considering that the operating data of a degradation level may have multiple forms of expression, there may be multiple attribute vectors for a certain degradation level. A is used to represent the set of all attribute vectors of all degradation levels. The compressor of the power plant data set used in this embodiment includes four attributes: glass fiber dirt, coarse filter dirt, fine filter dirt, and blade dirt. The attributes of the degradation level are measured using four evaluation standards: "excellent", "good", "medium", and "poor". The degradation level classification and attribute representation of the compressor are shown in Table 1.

[0053] Table 1: Degradation level classification and attribute representation of compressor in power plant dataset

[0054]

[0055] As shown in Table 1, the degradation process of the compressor is divided into five degradation levels, among which level 1, level 3 and level 4 correspond to two attribute vectors respectively, and level 2 and level 5 correspond to one attribute vector respectively. In addition, in order to quantify the values ​​of the evaluation criteria for the convenience of modeling, in the actual modeling process, the four evaluation criteria of "excellent", "good", "medium" and "poor" are mapped to the four values ​​of "0.25", "0.5", "0.75" and "1.00" respectively.

[0056] In an actual gas turbine system, the load is directly reflected in the generated power, which is controlled by the closed loop on the grid side and is a reflection of the compressor operating condition. Therefore, in this embodiment, the power is divided into Q load conditions, and each compressor operating data has a label corresponding to it in the load condition set W = {1, ..., Q}.

[0057] As mentioned above, in the actual power plant operating environment, there is a lack of data on high degradation levels. Therefore, this embodiment only uses operating data with low degradation levels (i.e., seen degradation levels) as the training set to train the evaluation model, so that it can correctly judge the degradation level when encountering new operating data. In this embodiment, a 60%-40% partitioning strategy is used to divide the multiple degradation levels into seen degradation levels and unseen degradation levels. That is, after all degradation levels are sorted from small to large according to the degree of performance degradation, the first 60% of the degradation levels are selected as seen degradation levels, and the remaining degradation levels are selected as unseen degradation levels. The training set is represented as in The operating data used for training indicates the degradation level seen during compressor operation, d x Indicates the data dimension, represents the set of running data used for training with the observed degradation level; y tr Represents x tr The corresponding degradation level label, represents the set of labels of the seen degradation levels, K s Indicates the number of degradation levels seen; Represents x tr The corresponding attribute vector, A represents the set of all attribute vectors of all degradation levels; w s Represents x tr The corresponding load condition labels, W = {1, ..., Q} represents the load condition label set, and Q represents the number of load conditions. Similarly, the test set is represented as where x te Indicates the running data to be tested. represents the set of operational data for which no degradation level is observed, The set of operational data used for the test that represents the degradation level seen; y te Represents x te The corresponding degradation level label, represents the set of unseen degradation level labels, K is the total number of all degradation levels; a u Represents x te The corresponding attribute vector; w u Represents x te In the traditional zero-shot learning setting for compressor performance degradation, it is necessary to train a classifier on the test set. That is, the classifier is able to evaluate samples from the seen degradation levels, while in the generalized zero-shot setting, it is necessary to train the classifier on the test set That is, the classifier can simultaneously evaluate samples from the seen degradation level and the unseen degradation level. Exemplarily, in this embodiment, the selected compressor comes from a certain type of gas turbine in a real thermal power plant. The power plant has installed multiple important sensor measurement points and data acquisition systems in the previous transformation, and the sampling frequency is 2 minutes. As shown in Table 1, the power plant data set selected in this embodiment contains 5 degradation levels, with a total of 8 attribute vectors. For these 5 degradation levels, a 60%-40% partitioning strategy is used to divide the seen degradation level and the unseen degradation level, that is, degradation levels 1, 2 and 3 are regarded as the seen degradation levels, and the remaining 2 are regarded as the unseen degradation levels; in addition, since this embodiment adopts the generalized zero sample setting, the training set contains 3 types of degradation levels, and the test data set contains all 5 types of degradation levels. At the same time, according to the actual operating conditions, the active power of the compressor is divided into 5 load conditions, that is, W = {1, ..., 5}, and each type of attribute description of each load condition in the data set contains 200 samples. Each sample contains 19 measurement points related to compressor performance degradation, such as compressor inlet and outlet temperatures, compressor bearing vibration, etc. The specific measurement point parameters, corresponding symbols and units are shown in Table 2.

[0058] Table 2: Measurement point names and units

[0059]

[0060]

[0061] (2) Constructing a domain adversarial structure network, which includes a feature extractor, a gradient reversal layer (GRL), a domain adversarial discriminator and an attribute predictor; using the training set to train the domain adversarial structure network to obtain the trained feature extractor, domain adversarial discriminator and attribute predictor.

[0062] In this embodiment, different load conditions are regarded as different domains, that is, each load condition is regarded as a domain, and a domain adversarial structure network is designed to extract consistent features under different load conditions, that is, domain invariant features. The gradient of the GRL input is multiplied by a negative number, thereby reversing the training objectives of the domain adversarial structure network before and after GRL to achieve the effect of adversarial training. Specifically, after the introduction of GRL, the optimization goal of the feature extractor will be opposite to the optimization goal of the domain adversarial discriminator, that is, the goal of the feature extractor is to extract features and make it as difficult as possible for the domain adversarial discriminator to determine which domain the feature comes from, while the goal of the domain adversarial discriminator is to determine the domain label of the correct feature as much as possible, and the two compete with each other until the domain adversarial structure network reaches a stable state. The domain adversarial discriminator enables the feature extractor to extract domain invariant features through the above-mentioned domain adversarial method, eliminating the problem of inconsistent correspondence between the measurement point values ​​and the compressor performance under different load conditions, and providing support for subsequent modeling. At the same time, based on domain adversarial learning, an attribute predictor is introduced to assist the feature extractor in extracting features with attribute differences and retain the differences between features of different attributes as much as possible to assist subsequent attribute-related generation tasks.

[0063] like Figure 2 As shown in the figure, the operating data under each load condition in the training set is input into the domain adversarial structure network. First, the features of the operating data are extracted by the feature extractor. The features extracted by the feature extractor are then passed through the gradient reversal layer and the domain adversarial discriminator in turn to obtain the probabilities corresponding to each domain label, and the domain label corresponding to the maximum probability is selected as the predicted domain label output by the domain adversarial discriminator. The features extracted by the feature extractor are sent to the attribute predictor to obtain the corresponding attribute vector prediction value. The cross entropy loss function is used to calculate the domain discrimination loss function based on the predicted domain label of the operating data and the corresponding true domain label in the training set. The binary cross entropy loss function is used to calculate the first label discrimination loss function based on the attribute vector prediction value output by the attribute predictor and the corresponding attribute vector true value in the training set. The domain discrimination loss function and the first label discrimination loss function are weighted to obtain the total loss function of the domain adversarial structure network. Taking minimizing the total loss function of the domain adversarial structure network as the optimization goal, the parameters of the feature extractor, domain adversarial discriminator and attribute predictor are adjusted to obtain the trained feature extractor, domain adversarial discriminator and attribute predictor. Among them, the trained feature extractor can extract domain invariant features with attribute differences.

[0064] Furthermore, the calculation formula of the domain discrimination loss function is:

[0065]

[0066] Among them, L D represents the domain discrimination loss function, f(·) represents the features output by the feature extractor; D a (·) represents the predicted domain label output by the domain adversarial discriminator, w s represents the true domain label and also the load condition label. P(·) represents the probability that the predicted domain label is the true domain label. It represents the mathematical expectation and the average loss on the training set.

[0067] Furthermore, the calculation formula of the first label discrimination loss function is:

[0068]

[0069] Among them, L A represents the first label discrimination loss function, M represents the total number of attributes in the attribute vector, and A(·) j represents the predicted value of the jth attribute output by the attribute predictor, a s,j Represents data x tr The true value of the jth attribute in the corresponding attribute vector.

[0070] (3) Constructing a Differential Generative Adversarial Network (DFGAN), which includes the trained feature extractor and attribute predictor, generator, converter, and generative adversarial discriminator obtained in step (2); using the training set to train the degraded differential generative adversarial network to obtain the trained generator, converter, and generative adversarial discriminator.

[0071] In this embodiment, based on the trained feature extractor and attribute predictor obtained in step (2), the structural design of the degradation difference generative adversarial network is described. Considering that the performance degradation process of the compressor has a progressive characteristic that becomes more serious with time, the degradation difference generative adversarial network learns the relationship between the semantic description and the real feature distribution. With the assistance of the attribute vector, it generates features with severe performance degradation (i.e., high-level features) through features with less performance degradation (i.e., low-level features), thereby converting the generalized zero-shot learning problem into a supervised learning problem. The high-level features are generated by using attributes and low-level features, which can solve the problem of difficulty in obtaining high-level training samples. Among them, "high level" represents the degradation level label y in the training set. tr The sample with a larger value is the opposite for the sample with a “lower level”. The structure of the degenerate difference generative adversarial network is as follows Figure 3As shown, the generator generates feature differences using the attribute differences between high-level attributes and low-level attributes, the converter is used to concatenate the feature differences with the low-level features and convert them into high-level features, and the generative adversarial discriminator enables the generator and converter to generate more realistic features in an adversarial manner.

[0072] Furthermore, in the process of training the degradation difference generative adversarial network using the training set, the operating data of the degradation levels of various load conditions in the training set of step (1) are input into the trained feature extractor obtained in step (2) to obtain domain-invariant degradation level features. Figure 3 As shown, a generative adversarial discriminator is introduced to train the generator and the converter. The training goal is to make the generator and the converter compete with the generative adversarial discriminator in a minimax game. The generative adversarial discriminator attempts to accurately distinguish real samples from generated samples, while the generator attempts to deceive the generative adversarial discriminator by generating samples that can be mistaken for real samples. The domain-invariant seen degradation level features extracted by the feature extractor are input to the generator and converter in the degraded difference generative adversarial network, and the lower-level features and attributes are used to generate higher-level features through the generator and converter. The features are then input into the generative adversarial discriminator and the trained attribute predictor obtained in step (2) and the corresponding loss function is constructed according to its output for training. After the training is completed, the trained generator and converter can be obtained. Among them, the optimization objective of the training corresponding to the loss function constructed according to the output of the generative adversarial discriminator is expressed as In addition, introducing the attribute predictor trained in the domain adversarial structure network into the degraded difference generative adversarial network can assist the generator in generating more attribute discriminative features.

[0073] Specifically, Figure 3 As shown, first randomly sample all samples in the training set and take out two samples Constitute a sample pair, where And assume because Value greater than value, so the sample For high-level samples, samples is a low-level sample; then calculate the high-level sample The attribute vector With low-level samples The attribute vector The difference between them is taken as the attribute difference, and then the attribute difference is spliced ​​with random noise and input into the generator to obtain the generated feature difference; at the same time, the high-level samples are Operation data and low-level samples Operation data Input it into the trained feature extractor to obtain the corresponding high-level features and low-level features, and use them as the true high-level features and true low-level features respectively; secondly, add the generated feature difference to the true low-level features and input them into the converter to obtain the generated high-level features, which can be expressed as: where G(·) represents the output of the generator, T(·) represents the output of the converter, and z represents random Gaussian noise. represents the generated high-level features; then, the real high-level features and the generated high-level features are respectively compared with the attributes of the high-level samples.

[0074] inside After being combined, they are input into the generative adversarial discriminator to obtain the discrimination result of the high-level features, that is, whether the high-level features input into the generative adversarial discriminator are real or generated; at the same time, the generated high-level features are input into the trained attribute predictor to obtain the corresponding attribute vector prediction value. The generative adversarial loss function is calculated based on the discrimination results corresponding to the real high-level features output by the feature extractor and the generated high-level features output by the converter. Using the binary cross entropy loss function, the attribute vector prediction value corresponding to the high-level features output by the attribute predictor and the attribute vector of the corresponding high-level samples in the training set are The real value is used to calculate the second label discrimination loss function. The generated adversarial loss function and the second label discrimination loss function are weighted to obtain the total loss function of the degraded difference generated adversarial network, and the optimization goal is to minimize the total loss function of the degraded difference generated adversarial network. Here, since the feature extractor and attribute predictor in the degraded difference generated adversarial network are trained in step (2), the parameters of the feature extractor and attribute predictor are frozen, and only the parameters of the generator, converter and generated adversarial discriminator need to be adjusted to obtain the trained generator, converter and generated adversarial discriminator. Among them, the trained generator and converter can use attribute vectors and low-level features to generate high-level features with attribute discrimination.

[0075] Furthermore, the calculation formula for generating the adversarial loss function is:

[0076]

[0077] in, represents the generated adversarial loss function, D b (·) represents the discriminant result of generating high-level features output by the adversarial discriminator, represents the high-level features output by the trained feature extractor, represents the high-level features generated by the converter output, λ represents the penalty coefficient, α~U(0,1) represents the weight coefficient, U(0,1) represents the uniform distribution defined on the interval [0,1], ||·||2 represents the second norm, express about The gradient of , E represents the mathematical expectation, which represents the average loss on all the sample pairs sampled above.

[0078] It should be understood that The first two terms in approximate the Wasserstein distance, and the third term is a gradient penalty that forces the gradient of the generative adversarial discriminator to have unit norm along the straight line between the real and generated sample pairs.

[0079] Furthermore, the calculation formula of the second label discrimination loss function is:

[0080]

[0081] Among them, L A ′ represents the second label discrimination loss function, Indicates high-level samples Running data in The corresponding attribute vector The true value of the jth attribute in .

[0082] (4) Based on the training set and the test set, the trained feature extractor obtained in step (2) is used to extract domain-invariant seen degradation level features. With the assistance of the attribute vector, the trained generator and transformer obtained in step (3) are used to generate domain-invariant unseen degradation level features. The classifier is supervisedly trained using the domain-invariant seen degradation level features and the domain-invariant unseen degradation level features and their corresponding degradation level labels to obtain a trained classifier for all degradation levels.

[0083] In this embodiment, a trained feature extractor can be obtained through the domain adversarial stage of step (2), and a trained generator and converter can be obtained through the DFGAN training stage of step (3). Generalized zero-shot learning intends to generate high-level features using attribute vector information and operating data in the training set when high-level degradation samples are missing, and train a classifier using the training set features and the generated features to achieve discrimination of all degradation levels. After the training is completed, a trained classifier for all degradation levels is finally obtained.

[0084] Specifically, Figure 4 As shown, first, the training set data x is sampled in the training set tr and its corresponding degradation level label y tr and attribute vector a s , randomly select attribute vector a from the unseen degradation levelsu and its corresponding degradation level label y u ; Then use the trained feature extractor to extract the training set data x tr The domain-invariant degradation level features seen are used as training set features; secondly, the unseen degradation level attribute vector a is calculated u The attribute vector a of the training set data s The attribute difference between them is then concatenated with random noise and input into the trained generator to obtain the generated feature difference; the generated feature difference is then added to the training set feature and input into the trained converter to obtain the generated unseen degradation level feature; finally, the generated unseen degradation level feature and the training set feature are respectively input into the classifier to obtain the corresponding predicted degradation level label. Using the cross entropy loss function, according to the unseen degradation level y u The generated features and corresponding predicted degradation level labels, as well as the predicted degradation level labels corresponding to the training set features and the degradation level labels corresponding to the training set data y tr , calculate the supervised training loss function of the classifier. Taking minimizing the supervised training loss function of the classifier as the optimization goal, adjust the parameters of the classifier to obtain a trained classifier. Among them, the trained classifier can simultaneously classify the features of the seen degradation level and the features of the unseen degradation level.

[0085] Furthermore, the calculation formula of the supervised training loss function of the classifier is:

[0086]

[0087] Among them, L C represents the supervised training loss function of the classifier, C(·) represents the predicted degradation level label output by the classifier, G(·) represents the output of the generator, T(·) represents the output of the transformer, and z represents random Gaussian noise.

[0088] (5) When applying, the operating data to be tested in the test set is first input into the trained feature extractor to obtain the corresponding domain-invariant features, and then the domain-invariant features are input into the trained classifier to obtain the corresponding degradation level. The higher the degradation level obtained, the more serious the compressor performance degradation is, that is, the more necessary it is to arrange water washing.

[0089] Specifically, the running data to be tested in the test set is input into the trained feature extractor, and its output is used as the input of the trained classifier. In order to alleviate the widespread domain shift problem in generalized zero-shot, that is, the model trained only with seen category data has poor generalization ability for unseen categories, a calibrated stacking method is used for the classifier, namely:

[0090]

[0091] in, is the running data x to be tested in the test set te The degradation level evaluation result of Indicates whether category c is a seen degradation level, and γ is the calibration coefficient.

[0092] The following uses the grade evaluation results of the test set of the embodiment to illustrate the effectiveness of the method of the present invention, that is, the zero-sample / generalized zero-sample evaluation accuracy, the generalized zero-sample seen and unseen evaluation accuracy and their harmonic mean to measure the effect of the method of the present invention. e Prediction, direct attribute prediction), IAP (Indirect Attribute Prediction), ESZSL (Embarrassingly Simple Zero-Shot Learning), Devise (Deep visual-semantic embedding model), ALE (Attribute Label Embedding), SJE (Structured Joint Embedding) are used for experimental comparison. In the DAP and IAP methods, replacing the base learner can obtain different effects. Here, Multilayer Perceptron (MLP), Ridge Regression (Ridge Regression) and Adaptive Boosting (AdaBoost) are used as base learners, and the results are shown in Table 3. Among them, the zero-shot learning (ZSL) method only evaluates the degradation assessment accuracy (ACC), while the generalized zero-shot learning (GZSL) method also evaluates the seen degradation level assessment accuracy (U), the unseen degradation level assessment accuracy (S), and the degradation level assessment harmonic average accuracy (H).

[0093] Table 3: Comparison of experimental results of different methods

[0094]

[0095] From the comparison results shown in Table 3, it can be seen that the zero-sample / generalized zero-sample evaluation accuracy of the method of the present invention is 0.9513 and 0.6460 respectively, and the generalized zero-sample seen and unseen evaluation accuracy and their harmonic means are 0.5207, 0.7400 and 0.6113 respectively. It can be found that the method of the present invention has a greater improvement than other methods in these five accuracy judgment indicators. The above proves the feasibility and effectiveness of the method proposed in the present invention.

[0096] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A zero-sample performance degradation assessment method for compressor water washing intelligent decision-making, characterized in that: The following steps are involved: (1) The performance degradation process of the gas turbine compressor is divided into multiple degradation levels, and the operating data of all degradation levels under various load conditions of the compressor are collected. The training set is constructed using the operating data of the degradation level that has been seen and the degradation level label, load condition label and attribute vector describing the degradation level corresponding to the operating data. The test set is constructed using the operating data of the degradation level that has not been seen and the operating data of the degradation level that has been seen and the degradation level label, load condition label and attribute vector describing the degradation level corresponding to the operating data. (2) constructing a domain adversarial structure network, which includes a feature extractor, a gradient reversal layer, a domain adversarial discriminator, and an attribute predictor; using a training set to train the domain adversarial structure network to obtain a trained feature extractor, a domain adversarial discriminator, and an attribute predictor; (3) constructing a degraded difference generative adversarial network, which includes the trained feature extractor and attribute predictor, generator, converter, and generative adversarial discriminator obtained in step (2); training the degraded difference generative adversarial network using the training set to obtain the trained generator, converter, and generative adversarial discriminator; (4) Based on the training set and the test set, the trained feature extractor obtained in step (2) is used to extract domain-invariant seen degradation level features, and the trained generator and transformer obtained in step (3) are used with the assistance of the attribute vector to generate domain-invariant unseen degradation level features. The domain-invariant seen degradation level features and the domain-invariant unseen degradation level features and their corresponding degradation level labels are used to perform supervised training on the classifier to obtain a trained classifier for all degradation levels; (5) When applying, the running data to be tested in the test set is first input into the trained feature extractor to obtain the corresponding domain-invariant features, and then the domain-invariant features are input into the trained classifier to obtain the corresponding degradation level.

2. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 1 is characterized in that: The observed degradation level and the unobserved degradation level are obtained by the following method: A 60%-40% partitioning strategy is adopted to divide the multiple degradation levels into seen degradation levels and unseen degradation levels. That is, after all degradation levels are sorted from small to large according to the degree of performance degradation, the first 60% of the degradation levels are selected as seen degradation levels, and the remaining degradation levels are selected as unseen degradation levels.

3. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 1 is characterized in that: The training set is represented as in, The operating data used for training indicates the degradation level seen during compressor operation, d x Indicates the data dimension, represents the set of running data used for training with the observed degradation level; y tr Represents x tr The corresponding degradation level label, represents the set of labels of the seen degradation levels, K s Indicates the number of degradation levels seen; Represents x tr The corresponding attribute vector, A represents the set of all attribute vectors of all degradation levels; w s Represents x tr The corresponding load condition labels, W = {1,…,Q} represents the set of load condition labels, and Q represents the number of load conditions; The test set is represented as Among them, x te Indicates the running data to be tested. represents the set of operational data for which no degradation level is observed, The set of operational data used for the test that represents the degradation level seen; y te Represents x te The corresponding degradation level label, represents the set of unseen degradation level labels, K is the total number of all degradation levels; a u Represents x te The corresponding attribute vector; w u Represents x te Corresponding load case label.

4. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 1 is characterized in that: The domain adversarial structure network is trained using the training set to obtain a trained feature extractor, a domain adversarial discriminator, and an attribute predictor, specifically including: The operating data under each load condition in the training set is input into the domain adversarial structure network. The features of the operating data are first extracted by the feature extractor. The features extracted by the feature extractor are then passed through the gradient reversal layer and the domain adversarial discriminator in turn to obtain the probabilities corresponding to each domain label. The domain label corresponding to the maximum probability is selected as the predicted domain label output by the domain adversarial discriminator. The features extracted by the feature extractor are sent to the attribute predictor to obtain the corresponding attribute vector prediction value. The cross entropy loss function is used to calculate the domain discrimination loss function according to the predicted domain labels of the running data and the corresponding true domain labels in the training set; the binary cross entropy loss function is used to calculate the first label discrimination loss function according to the predicted value of the attribute vector output by the attribute predictor and the corresponding true value of the attribute vector in the training set; the domain discrimination loss function and the first label discrimination loss function are weighted to obtain the total loss function of the domain adversarial structure network, and the parameters of the feature extractor, domain adversarial discriminator and attribute predictor are adjusted with the minimization of the total loss function of the domain adversarial structure network as the optimization goal to obtain the trained feature extractor, domain adversarial discriminator and attribute predictor; wherein the trained feature extractor can extract domain invariant features with attribute differences.

5. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 4 is characterized in that: The calculation formula of the domain discrimination loss function is: Among them, L D represents the domain discrimination loss function, f(·) represents the features output by the feature extractor, and D a (·) represents the predicted domain label output by the domain adversarial discriminator, w s represents the true domain label and also the load condition label. P(·) represents the probability that the predicted domain label is the true domain label. It represents the mathematical expectation and the average loss on the training set; The calculation formula of the first label discrimination loss function is: Among them, L A represents the first label discrimination loss function, M represents the total number of attributes in the attribute vector, and A(·) j represents the predicted value of the jth attribute output by the attribute predictor, a s,j Represents data x tr The true value of the jth attribute in the corresponding attribute vector.

6. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 1 is characterized in that: The method uses the training set to train the degraded difference generative adversarial network to obtain a trained generator, a converter, and a generative adversarial discriminator, specifically including: First, randomly sample all samples in the training set and take out two samples Constitute a sample pair, where And assume Then calculate the high-level samples The attribute vector With low-level samples The attribute vector The difference between them is taken as the attribute difference, and then the attribute difference is spliced ​​with random noise and input into the generator to obtain the generated feature difference; at the same time, the high-level samples are Operation data and low-level samples Operation data Input it into the trained feature extractor to obtain the corresponding high-level features and low-level features, and use them as the true high-level features and true low-level features respectively; secondly, add the generated feature difference to the true low-level features and input them into the converter to obtain the generated high-level features; then, add the true high-level features and the generated high-level features to the attribute vector of the high-level sample respectively. After the combination, the high-level features are input into the generative adversarial discriminator to obtain the discrimination result of the high-level features, that is, whether the high-level features input into the generative adversarial discriminator are real or generated; at the same time, the generated high-level features are input into the trained attribute predictor to obtain the corresponding attribute vector prediction value; The adversarial loss function is calculated based on the discrimination results corresponding to the real high-level features output by the feature extractor and the generated high-level features output by the converter; the binary cross entropy loss function is used to calculate the attribute vector prediction value corresponding to the high-level features output by the attribute predictor and the attribute vector of the corresponding high-level samples in the training set. The true value is used to calculate the second label discrimination loss function; the generative adversarial loss function and the second label discrimination loss function are weighted to obtain the total loss function of the degraded difference generative adversarial network, and the total loss function of the degraded difference generative adversarial network is minimized as the optimization goal, and the parameters of the generator, converter and generative adversarial discriminator are adjusted to obtain the trained generator, converter and generative adversarial discriminator; wherein the trained generator and converter can use attribute vectors and low-level features to generate high-level features with attribute discrimination.

7. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 6 is characterized in that: The calculation formula of the generated adversarial loss function is: in, represents the generated adversarial loss function, D b (·) represents the discriminant result of generating high-level features output by the adversarial discriminator, represents the high-level features output by the trained feature extractor, represents the high-level features generated by the converter output, λ represents the penalty coefficient, α~U(0,1) represents the weight coefficient, U(0,1) represents the uniform distribution defined on the interval [0,1], ∥·∥2 represents the second norm, express about The gradient of , E represents the mathematical expectation, which represents the average loss on all the sample pairs sampled above; The calculation formula of the second label discrimination loss function is: Among them, L A ′ represents the second label discrimination loss function, Indicates high-level samples Running data in The corresponding attribute vector The true value of the jth attribute in .

8. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 1 is characterized in that: The step (4) specifically comprises: First, sample the training set data x in the training set tr and its corresponding degradation level label y tr and attribute vector a s , randomly select attribute vector a from the unseen degradation levels u and its corresponding degradation level label y u ; Then use the trained feature extractor to extract the training set data x tr The domain-invariant degradation level features seen are used as training set features; secondly, the unseen degradation level attribute vector a is calculated u The attribute vector a of the training set data s The attribute difference between the two sets is obtained by splicing the attribute difference with random noise and then inputting it into the trained generator to obtain the generated feature difference; the generated feature difference is then added to the training set feature and inputted into the trained converter to obtain the generated unseen degradation level feature; finally, the generated unseen degradation level feature and the training set feature are respectively inputted into the classifier to obtain the corresponding predicted degradation level label; Using the cross entropy loss function, according to the unseen degradation level y u The generated features and corresponding predicted degradation level labels, as well as the predicted degradation level labels corresponding to the training set features and the degradation level labels corresponding to the training set data y tr , calculate the supervised training loss function of the classifier; take minimizing the supervised training loss function of the classifier as the optimization goal, adjust the parameters of the classifier, and obtain a trained classifier; wherein the trained classifier can simultaneously classify the seen degradation level features and the unseen degradation level features.

9. The zero-sample performance degradation assessment method for compressor water washing intelligent decision-making according to claim 8 is characterized in that: The calculation formula of the supervised training loss function of the classifier is: Among them, L C represents the supervised training loss function of the classifier, C(·) represents the predicted degradation level label output by the classifier, G(·) represents the output of the generator, T(·) represents the output of the transformer, and z represents random Gaussian noise.

Citation Information

Patent Citations

  • Engine gas path fault diagnosis method based on domain adversarial neural network

    CN115114977A

  • High-precision electronic equipment service life evaluation method combining physical information and generative adversarial network

    CN118396026A

  • Multimodal three-dimensional medical image fusion method and system, and electronic device

    WO2021022752A1

  • Multi-scale hybrid attention mechanism modeling method for predicting remaining useful life of aero engine

    WO2024087128A1