Zero-shot performance degradation assessment method for compressor water wash intelligent decision

Abstract

The application discloses a kind of zero-sample performance degradation evaluation methods for compressor water washing intelligent decision-making, the method designs compressor degradation knowledge-guided degradation difference generation network, with the aid of prior knowledge, the data characteristics of slight performance degradation are implemented directional degradation, generate the data characteristics of serious performance degradation, to train the classifier supervisedly, and according to the output degradation level, judge whether water washing;When the performance of compressor improves with the rise of gas turbine load, it will interfere with the evaluation of the true degradation degree of compressor, different load conditions are regarded as multiple domains, the features unchanged under each load condition are extracted from the operating data to alleviate the interference caused by load change, and the feature extractor retains the knowledge information in these features, improves the generation quality of degradation difference generation network.The application can evaluate the degradation level under the condition that the training data of serious degradation of compressor is missing, thereby providing support for the intelligent decision-making of gas generator set water washing.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment performance degradation evaluation, and particularly relates to a zero-sample performance degradation evaluation method for compressor water washing intelligent decision. BACKGROUND

[0002] Gas-steam combined cycle generator sets are widely used worldwide for their high efficiency and low pollution characteristics, and the gas-steam combined cycle generator sets 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 state of the entire power plant. As the core component of the gas turbine power generation system, the compressor is responsible for compressing air and gas for combustion. However, due to the inability of the air inlet filter system to completely remove all impurities in the air, the accumulated dirt on the compressor blades during long-term operation can reduce its efficiency, increase operating costs, and sometimes even cause surging, causing serious damage to the equipment. Therefore, regular off-line water washing of the compressor is crucial to ensure the safety, reliability and environmental protection of the power generation process.

[0003] At present, many power plants still determine the water washing period of the compressor according to the running time of the unit, but this fixed time interval cannot accurately reflect whether the actual performance of the compressor has decreased, thereby causing waste of water washing resources due to frequent water washing or causing the efficiency of the compressor to decrease due to untimely water washing. Therefore, more and more researchers begin to use a data-driven method to study the performance degradation degree of the compressor to assist in water washing decision-making. For example, some researchers propose a compressor reference pressure ratio model to obtain a reference pressure ratio interval and calculate the upper and lower limits of the pressure ratio drop. The greater the pressure ratio drop, the higher the degree of fouling of the compressor is considered. When the pressure ratio drop exceeds the water washing threshold, the gas turbine should be arranged for offline water washing. Some researchers select a characteristic index reflecting the degree of fouling of the compressor based on mechanism analysis, establish a reference value model of the characteristic index by comprehensively considering the influence of boundary conditions and operation methods, and make offline water washing decision-making based on the degree of fouling calculated based on the characteristic index in combination with a prior threshold. Some researchers use the deviation of the actual operating value of four performance parameters from the neural network predicted value under ideal conditions to measure the performance degradation degree of the compressor to provide a criterion for offline water washing. However, the existing data-driven methods use the thresholding method of performance evaluation indexes, but the performance of the compressor has a very high correlation with the power generation load. Under the operating condition of the highest power, the thermodynamic cycle of the gas turbine is closer to the ideal thermodynamic cycle, and at this time, the performance of the compressor is higher than that under the same conditions of other low-power operating conditions. Therefore, in the scene of frequent load changes of the gas turbine, the degradation level threshold of the index will also change. Secondly, the measured point value of the compressor changes with the load change controlled by the power grid side, so the model needs to consider the corresponding relationship between the measured point value and the performance of the compressor under different load conditions. In addition, the power plant usually does not allow the compressor to operate to a serious degradation degree, so the data of high degradation degree is difficult to obtain, and the corresponding degradation threshold cannot be set according to the data.

[0004] In view of the problems that the threshold is difficult to set due to the frequent load change characteristics of the gas turbine and the lack of data of high degradation degree, an expert knowledge assisted zero sample learning strategy is needed to evaluate the performance degradation degree of the compressor. SUMMARY

[0005] The present application aims at the deficiencies of the prior art, and provides a zero sample performance degradation evaluation method for intelligent decision-making of compressor water washing. The present application provides a solution to the problem that the performance degradation level evaluation model cannot be established due to the unobtainable data of serious degradation, and improves the accuracy of zero sample performance degradation evaluation in industrial processes.

[0006] The purpose of the present application is achieved by the following technical scheme: a zero sample performance degradation evaluation method for intelligent decision-making of compressor water washing, comprising the following steps:

[0007] (1) dividing the performance degradation process of a gas turbine compressor into multiple degradation levels, collecting operation data of all degradation levels under various load conditions of the compressor, using the operation data of seen degradation levels for training and the degradation level labels, load condition labels and attribute vectors describing the degradation levels corresponding to the operation data to construct a training set, and using the operation data of unseen degradation levels and the operation data of seen degradation levels for testing and the degradation level labels, load condition labels and attribute vectors describing the degradation levels corresponding to the operation data to construct a test set;

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

[0009] (3) constructing a degradation difference generative adversarial network including the trained feature extractor and attribute predictor obtained in step (2), a generator, a converter and a generative adversarial discriminator, training the degradation 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 test set, extracting domain-invariant seen degradation level features using the trained feature extractor obtained in step (2), generating domain-invariant unseen degradation level features using the trained generator and converter obtained in step (3) with the aid of the attribute vectors, and performing supervised training on the classifier using the domain-invariant seen degradation level features, domain-invariant unseen degradation level features and their corresponding degradation level labels to obtain the trained classifier for all degradation levels;

[0011] (5) in application, first inputting the operation data to be tested in the test set into the trained feature extractor to obtain corresponding domain-invariant features, and then inputting the domain-invariant features into the trained classifier to obtain corresponding degradation levels.

[0012] Further, the seen degradation levels and the unseen degradation levels are obtained by the following method:

[0013] The multiple degradation levels are divided into seen degradation levels and unseen degradation levels using a 60%-40% division strategy, that is, after all the degradation levels are sorted in ascending order of performance degradation degree, the first 60% of the total degradation levels are selected as seen degradation levels, and the remaining degradation levels are selected as unseen degradation levels.

[0014] Further, the training set is represented as wherein, training data representing seen degradation levels during compressor operation, d x representing data dimensions, a set of training data representing seen degradation levels; y tr representing x tr corresponding degradation level labels, a set of seen degradation level labels, K s representing the number of seen degradation levels; representing x tr corresponding attribute vectors, A represents a set of all attribute vectors of all degradation levels; w s representing x tr corresponding load condition labels, W = {1,..., Q} represents a set of load condition labels, Q represents the number of load conditions;

[0015] The test set is represented as wherein x te representing the operation data to be tested, a set of operation data representing unseen degradation levels, a set of operation data representing seen degradation levels for testing; y te representing x te corresponding degradation level labels, a set of unseen degradation level labels, K is the total number of all degradation levels; a u representing x te corresponding attribute vectors; w u representing x te corresponding load condition labels.

[0016] Further, the training set is used to train the domain adversarial structure network to obtain a trained feature extractor, domain adversarial discriminator and attribute predictor, specifically comprising:

[0017] The operation data under each load condition in the training set is input into the domain adversarial structure network, and the feature of the operation data is extracted by the feature extractor first; the feature extracted by the feature extractor is sequentially subjected to the gradient reversal layer and the domain adversarial discriminator to obtain the probability 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 feature extracted by the feature extractor is sent to the attribute predictor to obtain the corresponding attribute vector prediction value;

[0018] The domain discrimination loss function is calculated according to the predicted domain label of the running data and the corresponding real domain label in the training set by using a cross-entropy loss function; the first label discrimination loss function is calculated according to the attribute vector predicted value output by the attribute predictor and the corresponding attribute vector real value in the training set by using a binary cross-entropy loss function; 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, the domain adversarial discriminator and the attribute predictor are adjusted to minimize the total loss function of the domain adversarial structure network as an optimization goal, so as 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] Further, the calculation formula of the domain discrimination loss function is:

[0020]

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

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

[0023]

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

[0025] Further, the training set is used to train the degradation difference generation adversarial network to obtain the trained generator, converter and generation adversarial discriminator, and the specific steps include:

[0026] First, all samples in the training set are randomly sampled to obtain two samples to form a sample pair, wherein and it is assumed that Then, the attribute vector of the high-level sample is calculated. attribute vector of the high-level sample the difference between the generated feature difference value and the real low-level feature as a second label discriminative loss function, and weighting the generated adversarial loss function and the second label discriminative loss function to obtain a total loss function of the degradation difference adversarial network. running data of the high-level sample and the running data of the low-level sample are input into the trained feature extractor to obtain corresponding high-level features and low-level features, which are respectively taken as real high-level features and real low-level features; secondly, the generated feature difference value and the real low-level feature are added and input into the converter to obtain generated high-level features; then, the real high-level features and the generated high-level features are combined with the attribute vector of the high-level sample and input into the trained generative adversarial discriminator to obtain a discriminative result of the high-level features, i.e., whether the high-level features input into the generative adversarial discriminator are real or generated; meanwhile, the generated high-level features are input into the trained attribute predictor to obtain a corresponding attribute vector prediction value.

[0027] The generated adversarial loss function is calculated according to the discriminative result of the real high-level features output by the feature extractor and the generated high-level features output by the converter; the second label discriminative loss function is calculated by using a binary cross-entropy loss function according to the attribute vector prediction value of the high-level features output by the attribute predictor and the attribute vector of the high-level sample in the training set ; the generated adversarial loss function and the second label discriminative loss function are weighted to obtain the total loss function of the degradation difference adversarial network, the parameters of the generator, the converter and the generative adversarial discriminator are adjusted to minimize the total loss function of the degradation difference adversarial network as an optimization target, and the trained generator, the converter and the generative adversarial discriminator are obtained; wherein the trained generator and the converter can generate high-level features with attribute discriminability by using the attribute vector and the low-level feature.

[0028] Further, the calculation formula of the generated adversarial loss function is as follows:

[0029]

[0030] wherein, D (·) represents the discriminative result of the high-level features output by the generative adversarial discriminator, b D (·) represents the discriminative result of the high-level features output by the generative adversarial discriminator, represents the high-level features output by the trained feature extractor, represents the generated high-level features output by the converter, and λ represents a penalty coefficient. ​wherein, a ~ U(0, 1) denotes a weight coefficient, U(0, 1) denotes a uniform distribution defined in the interval [0, 1], ||·||2 denotes a two-norm, Table Regarding the gradient of the loss function, E denotes a mathematical expectation, and represents an average loss on all sampled sample pairs;

[0031] The calculation formula of the second label discriminative loss function is as follows:

[0032]

[0033] wherein, L A ′ denotes the second label discriminative loss function, denotes a high-level sample running data corresponding attribute vector the true value of the jth attribute in the attribute vector.

[0034] Further, the step (4) specifically comprises:

[0035] First, sample training set data x tr and its corresponding degradation level label y tr and attribute vector a s in the training set, randomly select attribute vector a u and its corresponding degradation level label y u in the unseen degradation level; then extract the domain-invariant seen degradation level features of the training set data x tr using the trained feature extractor as the training set features; secondly, calculate the attribute difference between the unseen degradation level attribute vector a u and the attribute vector a s of the training set data, and then input the attribute difference concatenated with random noise into the trained generator to obtain the generated feature difference; then input the generated feature difference and the training set features into the trained converter to obtain the generated unseen degradation level features; finally, input the generated unseen degradation level features and the training set features into the classifier to obtain the corresponding predicted degradation level label;

[0036] According to the unseen degradation level y u and its generated features, the corresponding predicted degradation level label, and the predicted degradation level label corresponding to the training set features and the degradation level label y tr, calculate a supervised training loss function of the classifier; adjust parameters of the classifier to minimize the supervised training loss function of the classifier as an optimization goal, to obtain a trained classifier; wherein the trained classifier can classify seen degradation level features and unseen degradation level features.

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

[0038]

[0039] wherein, 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 converter, and z represents random Gaussian noise.

[0040] The beneficial effects of the present application are that the present application proposes a knowledge data fusion zero sample performance degradation evaluation method for compressor performance level evaluation problems; in addition, since the compressor performance and power generation load have very high correlation, in order to consider the influence of variable load on the judgment of compressor performance degradation degree, consistent prior knowledge under each load condition is designed for each degradation level to describe the compressor performance state, 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 the generative zero sample learning, the present application proposes a degradation difference generative adversarial network for the degradation zero sample problem, introduces the idea of performance gradual degradation, generates high degradation level features by using low degradation level features with the aid of knowledge, overcomes the problem that modeling is impossible due to difficulty in obtaining high degradation level data, and improves the quality of generated data; the present application can mine the gradual change characteristics of compressor performance degradation, improve the accuracy of zero sample performance degradation evaluation in industrial processes, and assist in compressor water washing decision. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The flowchart of the zero sample performance degradation evaluation method for the intelligent decision of the compressor water washing of the present application;

[0042] Figure 2 The structure diagram of the domain adversarial structure network of the present application;

[0043] Figure 3 The structure diagram of the degradation difference generative adversarial network of the present application;

[0044] Figure 4 The flowchart of training the classifier by using the feature extractor, the generator and the converter of the present application. DETAILED DESCRIPTION

[0045] The exemplary embodiments will be described in detail herein with reference to the attached drawings. Descriptions of well-known functions and constructions can be omitted to avoid obscuring the point of the present application. The following description is presented to enable any person skilled in the art to make and use the application. Descriptions of specific devices and applications are provided only as examples. Any person skilled in the art and others can readily devise other methods and applications that incorporate the various principles of the present application discussed below.

[0046] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0047] It is to be understood that the singular forms "a", "an", and "the" include plural referents unless the context clearly dictates otherwise. It is to be understood that the term "comprising" as used herein is to be taken as meaning "including, but not limited to". It is to be understood that the term "comprising" as used herein is to be taken as meaning "including, but not limited to". As used herein, the terms "if' and "as soon as" can each be interpreted to express the concept of "in response to a determination" or "when it is determined".

[0048] The present application will be described in detail below with reference to the drawings. The features of the embodiments and implementation described below can be combined with each other as long as there is no conflict.

[0049] The present embodiment verifies the effectiveness of the present application based on actual measurement data of a certain thermal power plant, but is not limited thereto. Any industrial engineering with progressive degradation characteristics can use the method described in the present application to evaluate the performance degradation level with zero samples.

[0050] Referring to Figure 1 The zero-sample performance degradation evaluation method for compressor water washing intelligent decision of the present application specifically includes the following steps:

[0051] (1) The performance degradation process of the 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 for testing of the seen degradation levels 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 influence of variable load on the compressor performance degradation threshold, expert prior knowledge is designed for each degradation level to reflect the compressor performance state without changing the load, wherein the prior knowledge includes M attributes. Experts use four evaluation criteria, i.e., "excellent", "good", "medium", and "poor", to measure the attributes of each degradation level. Of course, the selection of evaluation criteria can be selected according to actual needs, for example, seven more detailed evaluation criteria, i.e., "excellent up", "excellent down", "good up", "good down", "medium up", "medium down", and "poor", are used. A combination of evaluation criteria containing all attributes is called an attribute vector. Considering that the operating data of a degradation level can have multiple forms of expression, for a certain degradation level, there can be multiple attribute vectors. 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, i.e., glass fiber contamination, coarse filter contamination, fine filter contamination, and blade contamination. Four evaluation criteria, i.e., "excellent", "good", "medium", and "poor", are used to measure the attributes of the degradation level. The degradation level division of the compressor of the power plant data set and the attribute representation are shown in Table 1.

[0053] Table 1: Degradation level division of the compressor of the power plant data set and attribute representation

[0054]

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

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

[0057] As mentioned above, in the actual power plant operation environment, high degradation level data is missing, so this embodiment only uses low degradation level (i.e. seen degradation level) operation data as a training set to train the evaluation model, which can correctly judge the belonging degradation level when encountering new operation data. In this embodiment, 60%-40% division strategy is used to divide the multiple degradation levels into seen degradation level and unseen degradation level, that is, after all degradation levels are sorted according to the performance degradation degree from small to large, the top 60% of the total degradation levels are selected as seen degradation level, and the remaining degradation levels are selected as unseen degradation level. The training set is represented as wherein represents the operation data for training of seen degradation level in the process of compressor operation, d x represents the data dimension, represents the operation data set for training of seen degradation level; y tr represents x tr corresponding degradation level label, represents the seen degradation level label set, K s represents the number of seen degradation levels; represents x tr corresponding attribute vector, A represents the attribute vector set of all degradation levels; w s represents x tr corresponding load condition label, W = {1,..., Q} represents the load condition label set, and Q represents the number of load conditions. Similarly, the test set is represented as wherein x te represents the operation data to be tested, represents the operation data set of unseen degradation level, represents the operation data set for testing of seen degradation level; y te represents x te corresponding degradation level label, represents the unseen degradation level label set, K is the total number of all degradation levels; a u represents x te corresponding attribute vector; w u represents x te corresponding load condition label. In the traditional zero sample learning setting of compressor performance degradation, the classifier on the test set needs to be trained that is, the classifier can evaluate samples from seen degradation level, and in the generalized zero sample setting, the classifier on the test set needs to be trained That is, the classifier can evaluate samples from seen degradation levels and unseen degradation levels simultaneously. Exemplarily, in the present embodiment, the selected compressor is from a certain type of gas turbine in a real thermal power plant, and the power plant has installed a number of sensor measuring points and a data acquisition system in the important parts in the previous transformation, and the sampling frequency is 2 minutes. As shown in Table 1, the data set selected in the present embodiment contains 5 degradation levels, and a total of 8 attribute vectors. For the 5 degradation levels, a 60%-40% division strategy is used to divide the seen degradation levels and the unseen degradation levels, that is, degradation levels 1, 2 and 3 are taken as seen degradation levels, and the remaining 2 are taken as unseen degradation levels; in addition, since the present embodiment adopts the generalized zero sample setting, the training set contains 3 degradation level types, and the test data set contains all 5 degradation levels. At the same time, according to the actual operation, the compressor active power is divided into 5 load conditions, that is, W = {1,..., 5}, and each load condition in the data set contains 200 samples of each attribute description. Each sample contains 19 measuring points related to the performance degradation of the compressor, such as compressor inlet and outlet temperature, compressor bearing vibration, etc., and the specific measuring point parameters, corresponding symbols and units are shown in Table 2.

[0058] Table 2: Measuring point name and unit

[0059]

[0060]

[0061] (2) Constructing a domain adversarial structure network, the domain adversarial structure network comprising 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, the domain adversarial discriminator and the 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 target of the domain adversarial structure network before and after the GRL to achieve the effect of adversarial training. Specifically, when the GRL is introduced, the optimization target of the feature extractor will be opposite to that 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 judge the domain of the features, while the goal of the domain adversarial discriminator is to judge the domain label of the correct features as much as possible, and the two are in a game 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 domain adversarial method, eliminating the inconsistency of the corresponding relationship between the measured point values and the compressor performance under different load conditions, and providing support for subsequent modeling. At the same time, an attribute predictor is introduced on the basis of domain adversarial, which is used to assist the feature extractor to extract 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] As shown in Figure 2 The running data of each load condition in the training set is input into the domain adversarial structure network, and the features of the running data are first extracted by the feature extractor. The features extracted by the feature extractor are sequentially input into the gradient reversal layer and the domain adversarial discriminator to obtain the probability 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 input into the attribute predictor to obtain the corresponding attribute vector prediction value. The domain discrimination loss function is calculated according to the predicted domain label of the running data and the corresponding true domain label in the training set by using the cross entropy loss function. The first label discrimination loss function is calculated according to the attribute vector prediction value output by the attribute predictor and the corresponding attribute vector true value in the training set by using the binary cross entropy loss function. 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. The parameters of the feature extractor, the domain adversarial discriminator and the attribute predictor are adjusted to minimize the total loss function of the domain adversarial structure network as the optimization target, and the trained feature extractor, domain adversarial discriminator and attribute predictor are obtained. Among them, the trained feature extractor can extract domain-invariant features with attribute differences.

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

[0065]

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

[0067] Further, the calculation formula of the first label discriminative loss function is:

[0068]

[0069] where L A represents the first label discriminative loss function, M represents the total number of attributes in the attribute vector, A(·) j represents the predicted value of the jth attribute output by the attribute predictor, a s，j represents the real value of the jth attribute in the corresponding attribute vector. tr

[0070] (3) Construct a degradation difference generative adversarial network (DFGAN), which includes the trained feature extractor and attribute predictor obtained in step (2), a generator, a converter, and a generative adversarial discriminator; use the training set to train the degradation difference generative adversarial network to obtain a trained generator, converter, and generative adversarial discriminator.

[0071] In this embodiment, the structure design of the degradation difference generative adversarial network is described based on the trained feature extractor and attribute predictor obtained in step (2). Considering that the performance degradation process of the compressor has the progressive characteristic of becoming more and more serious over time, the degradation difference generative adversarial network learns the relationship between the semantic description and the real feature distribution, generates the performance degradation serious feature (i.e. high-level feature) through the performance degradation light feature (i.e. low-level feature) with the aid of the attribute vector, thereby converting the generalized zero-shot learning problem into a supervised learning problem, and generating the high-level feature using the attribute and low-level feature can solve the problem that the high-level training sample is difficult to obtain. Among them, “high level” represents the sample with a larger value of the degradation level label y tr in the training set, and “low level” is the opposite. The structure of the degradation difference generative adversarial network is as follows: Figure 3 ​As shown, the generator uses the attribute difference between high-level attributes and low-level attributes to generate feature difference values. The converter is used to concatenate the feature difference values ​​with low-level features and convert them into high-level features. The generative adversarial discriminator enables the generator and converter to generate more realistic features through adversarial methods.

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

[0073] Specifically, such as Figure 3 As shown, firstly, all samples in the training set are randomly sampled, and two samples are selected. Forming sample pairs, where And assume because Value greater than Value, therefore the sample This is a high-level sample. For low-level samples; then calculate high-level samples. attribute vector Compared with low-level samples attribute vector The difference between the samples is used as the attribute difference. This attribute difference is then concatenated with random noise and input into the generator to obtain the generated feature difference. Simultaneously, high-level samples are... running data and low-level samples runtime data The input is fed into the trained feature extractor to obtain the corresponding high-level and low-level features, which are then used as the true high-level and low-level features, respectively. Next, the generated feature difference is added to the true low-level features and then input into the converter to obtain the generated high-level features, which can be represented as: Where G(·) represents the generator output, T(·) represents the converter output, and z represents random Gaussian noise. This 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 The combined high-level features are then fed into a generative adversarial discriminator (GAD) to determine whether they are real or generated. Simultaneously, the generated high-level features are fed into a trained attribute predictor to obtain corresponding attribute vector predictions. The GAD loss function is calculated based on the discriminant results corresponding to the real high-level features output by the feature extractor and the generated high-level features output by the converter. A binary cross-entropy loss function is then used, based on the attribute vector predictions corresponding to the high-level features output by the attribute predictor and the attribute vectors of the corresponding high-level samples in the training set. The true value is used to calculate the second label discriminant loss function. The generative adversarial loss function and the second label discriminant loss function are weighted to obtain the total loss function of the degenerate difference generative adversarial network. The optimization objective is to minimize the total loss function of the degenerate difference generative adversarial network. Since the feature extractor and attribute predictor in the degenerate difference generative adversarial network are trained in step (2), the parameters of the feature extractor and attribute predictor are frozen. Only the parameters of the generator, converter and generative adversarial discriminator need to be adjusted to obtain the trained generator, converter and generative adversarial discriminator. Among them, the trained generator and converter can generate high-level features with attribute discriminative power using attribute vectors and low-level features.

[0075] Furthermore, the formula for calculating the adversarial loss function is as follows:

[0076]

[0077] in, D represents the generative adversarial loss function. b (·) represents the discrimination result of the high-level features output by the adversarial discriminator. This represents the high-level features output by the trained feature extractor. This represents the high-level characteristics generated by the converter output, where λ represents the penalty coefficient. where a U (0, 1) denotes the weight coefficient, U (0, 1) denotes the uniform distribution defined on the interval [0, 1], and ||·||2 denotes the two-norm, denotes the gradient of L , E denotes the mathematical expectation, and represents the average loss on all sampled sample pairs.

[0078] It should be understood that the first two terms in L

[0079] approximate the Wasserstein distance, and the third term is a gradient penalty that forces the gradient of the generated adversarial discriminator to have a unit norm along the straight line between the real and generated sample pairs.

[0080]

[0081] where L A denotes the second label discriminative loss function, denotes the high-level sample , the running data in L corresponds to the attribute vector , and the true value of the jth attribute in L

[0082] .

[0083] In this embodiment, the trained feature extractor can be obtained after the domain adversarial stage of step (2), and the trained generator and converter can be obtained after the DFGAN training stage of step (3). In the case of a lack of high-level degraded samples, the generalized zero-shot learning generates high-level features using attribute vector information and running data in the training set, trains a classifier using training set features and generated features, discriminates all degradation levels, and finally obtains a trained classifier for all degradation levels.

[0084] Specifically, as shown in FIG. 2, first, the training set data x tr , the corresponding degradation level label y tr , and the attribute vector a s are sampled in the training set, and the attribute vector a u is randomly selected in the unseen degradation level.u and its corresponding degradation level label y u ; then the trained feature extractor is used to extract the domain-invariant seen degradation level features of the training set data x tr , as the training set features; secondly, the attribute difference between the attribute vector a u of the training set data and the attribute vector a s of the training set data is calculated, and then the attribute difference is concatenated with random noise and input into the trained generator to obtain the generated feature difference; the generated feature difference is added to the training set features and input into the trained converter to obtain the generated unseen degradation level features; finally, the generated unseen degradation level features and the training set features are input into the classifier to obtain the corresponding predicted degradation level labels. Using the cross-entropy loss function, the supervised training loss function of the classifier is calculated according to the unseen degradation level y u , its generated features, and the corresponding predicted degradation level labels, as well as the predicted degradation level labels of the training set features and the degradation level labels y tr of the training set data. The supervised training loss function of the classifier is minimized as the optimization objective to adjust the parameters of the classifier to obtain the trained classifier. The trained classifier can classify both the seen degradation level features and the unseen degradation level features.

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

[0086]

[0087] wherein, 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 converter, and z represents random Gaussian noise.

[0088] (5) In application, the running data to be tested in the test set is 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 is, the more serious the degradation of the compressor performance is, that is, the more necessary water washing is arranged.

[0089] Specifically, the running data to be tested in the test set is input into the trained feature extractor, and the output thereof is used as the input of the trained classifier. To alleviate the domain shift problem widely existing in the generalized zero-shot learning, that is, the generalization ability of the model trained only using seen class data is poor for unseen classes, the calibrated stacking method is adopted for the classifier, that is:

[0090]

[0091] in, To test the runtime data x in the test set te The degradation level assessment results, indicator function Indicates whether category c is a known degradation level, and γ is the calibration coefficient.

[0092] The effectiveness of the method described in this invention is illustrated below using the ranking evaluation results of the example test set. Specifically, the effectiveness of the method is measured by the zero-sample / generalized zero-sample evaluation accuracy, the generalized zero-sample seen / unseen evaluation accuracy, and their harmonic mean. Direct Attribute Prediction (DAP) is selected here. e The following methods were used for experimental comparison: Direct Attribute Prediction (DAP), Indirect Attribute Prediction (IAP), Embarrassingly Simple Zero-Shot Learning (ESZSL), Devise (Deep visual-semantic embedding model), ALE (Attribute Label Embedding), and SJE (Structured Joint Embedding). Different results were obtained by replacing the base learner in the DAP and IAP methods. Multilayer Perceptron (MLP), Ridge Regression, and Adaptive Boosting (AdaBoost) were used as base learners, and the results are shown in Table 3. The Zero-Shot Learning (ZSL) method only evaluated the degradation assessment accuracy (ACC), while the Generalized Zero-Shot Learning (GZSL) method also evaluated the accuracy of seen degradation levels (U), the accuracy of unseen degradation levels (S), and the harmonic mean accuracy of degradation level assessment (H).

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

[0094]

[0095] From the comparison results shown in Table 3, it can be seen that the zero-shot / generic zero-shot evaluation accuracy of the method is 0.9513 and 0.6460 respectively, and the generic zero-shot seen, unseen evaluation accuracy and harmonic mean are 0.5207, 0.7400 and 0.6113 respectively. It can be found that the method has greater improvement than other methods in the five precision judgment indexes. The above proves the feasibility and effectiveness of the method.

[0096] The above examples are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A zero-sample performance degradation assessment method for intelligent decision-making in compressor water washing, characterized in that, Includes the following steps: (1) Divide the performance degradation process of the gas turbine compressor into multiple degradation levels, collect the operating data of all degradation levels under various load conditions of the compressor, use the operating data of known degradation levels for training, as well as the degradation level labels, load condition labels and attribute vectors describing the degradation levels corresponding to the operating data to construct a training set, and use the operating data of unknown degradation levels and the operating data of known degradation levels for testing, as well as the degradation level labels, load condition labels and attribute vectors describing the degradation levels corresponding to the operating data to construct a test set; (2) Construct a domain adversarial structure network, which includes a feature extractor, a gradient reversal layer, a domain adversarial discriminator, and an attribute predictor; train the domain adversarial structure network using a training set to obtain the trained feature extractor, domain adversarial discriminator, and attribute predictor. (3) Construct a degenerate difference generative adversarial network, which includes the trained feature extractor and attribute predictor, generator, converter and generative adversarial discriminator obtained in step (2); train the degenerate 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, use the trained feature extractor obtained in step (2) to extract domain-invariant known degradation level features. With the help of the attribute vector, use the trained generator and converter obtained in step (3) to generate domain-invariant unseen degradation level features. Use the domain-invariant known degradation level features and the domain-invariant unseen degradation level features and their corresponding degradation level labels to conduct supervised training on the classifier to obtain a trained classifier for all degradation levels. (5) When applying, first input the running data to be tested in the test set into the trained feature extractor to obtain the corresponding domain-invariant features, and then input the domain-invariant features into the trained classifier to obtain the corresponding degradation level.

2. The zero-sample performance degradation evaluation method for intelligent decision-making in compressor water washing according to claim 1, characterized in that, The observed degradation level and the unobserved degradation level are obtained by the following method: A 60%-40% segmentation strategy is adopted to divide the multiple degradation levels into seen degradation levels and unseen degradation levels. That is, after sorting all degradation levels according to the degree of performance degradation from smallest to largest, the top 60% of the total number of 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 evaluation method for intelligent decision-making in compressor water washing according to claim 1, characterized in that, The training set is represented as in, This indicates training data representing the level of degradation observed during compressor operation, d. x Indicates data dimensions, This represents the set of training data for which the degradation level has been observed; y tr x represents tr The corresponding degradation level label, This indicates that the set of degradation level labels has been seen, K. s Indicates the number of degradation levels that have been observed; x represents tr The corresponding attribute vector, A represents the set of all attribute vectors for all degradation levels; w s x represents tr The corresponding load condition label, W={1,…,Q} represents the set of load condition labels, and Q represents the number of load conditions; The test set is represented as Where, x te This represents the runtime data to be tested. This represents the set of operational data for which no degradation level was observed. This represents the set of runtime data used for testing that indicates the observed degradation level; y te x represents te The corresponding degradation level label, This indicates a set of no degradation level labels found, where K is the total number of all degradation levels; a u x represents te The corresponding attribute vector; w u x represents te The corresponding load condition label.

4. The zero-sample performance degradation evaluation method for intelligent decision-making in compressor water washing according to claim 1, characterized in that, The step of training the domain adversarial structure network using a training set to obtain a trained feature extractor, domain adversarial discriminator, and attribute predictor specifically includes: The operational data under various load conditions in the training set are input into the domain adversarial structure network. First, the feature extractor extracts the features of the operational data. The features extracted by the feature extractor are then passed through the gradient inversion layer and the domain adversarial discriminator to obtain the probability corresponding to each domain label. The domain label with the highest probability is selected as the predicted domain label output by the domain adversarial discriminator. The features extracted by the feature extractor are fed into the attribute predictor to obtain the corresponding attribute vector prediction value. Using the cross-entropy loss function, a domain discriminant loss function is calculated based on the predicted domain labels of the running data and the corresponding true domain labels in the training set. Using the binary cross-entropy loss function, a first label discriminant loss function is calculated based on the predicted attribute vector values ​​output by the attribute predictor and the corresponding true attribute vector values ​​in the training set. The domain discriminant loss function and the first label discriminant loss function are weighted to obtain the total loss function of the domain adversarial structure network. With minimizing the total loss function of the domain adversarial structure network as the optimization objective, the parameters of the feature extractor, the domain adversarial discriminator, and the attribute predictor are adjusted to obtain the trained feature extractor, the domain adversarial discriminator, and the attribute predictor. Among them, the trained feature extractor can extract domain-invariant features with attribute differences.

5. The zero-sample performance degradation evaluation method for intelligent decision-making in compressor water washing according to claim 4, characterized in that, The formula for calculating the domain discriminant loss function is as follows: Among them, L D Let f(·) represent the domain discriminant loss function, f(·) represent the features output by the feature extractor, and D represent the domain discriminant loss function. a (·) represents the predicted domain label output by the domain adversarial discriminator, w s This 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. This represents the mathematical expectation, which is the average loss on the training set. The formula for calculating the first label discrimination loss function is: Among them, L A Let A(·) represent the first label discriminant loss function, M represent the total number of attributes in the attribute vector, and A(·) represent the first label discriminant loss function. j a represents the predicted value of the j-th attribute output by the attribute predictor. s,j Representing data x tr The actual value of the j-th attribute in the corresponding attribute vector.

6. The zero-sample performance degradation evaluation method for intelligent decision-making in compressor water washing according to claim 1, characterized in that, The step of training the degradation interpolation generative adversarial network using a training set to obtain a trained generator, converter, and generative adversarial discriminator specifically includes: First, randomly sample all samples in the training set and select two samples. Forming sample pairs, where And assume Then calculate the high-level samples. attribute vector Compared with low-level samples attribute vector The difference between the samples is used as the attribute difference. This attribute difference is then concatenated with random noise and input into the generator to obtain the generated feature difference. Simultaneously, high-level samples are... runtime data and low-level samples runtime data The input is fed into a trained feature extractor to obtain corresponding high-level and low-level features, which are then used as the true high-level and low-level features, respectively. Next, the generated feature difference is added to the true low-level features and then input into a converter to obtain the generated high-level features. Finally, the true high-level features and the generated high-level features are respectively compared with the attribute vector of the high-level sample. The combined high-level features are then input into the generative adversarial discriminator to obtain the discrimination result of the high-level features, i.e., 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 true high-level features output by the feature extractor and the generated high-level features output by the converter. Then, using the binary cross-entropy loss function, the predicted attribute vectors corresponding to the high-level features output by the attribute predictor and the attribute vectors of the corresponding high-level samples in the training set are used. The second label discriminant loss function is calculated based on the true values. The generative adversarial loss function and the second label discriminant loss function are weighted to obtain the total loss function of the degradation difference generative adversarial network. The parameters of the generator, transformer, and generative adversarial discriminator are adjusted with the goal of minimizing the total loss function of the degradation difference generative adversarial network to obtain the trained generator, transformer, and generative adversarial discriminator. The trained generator and transformer can generate high-level features with attribute discriminative power using attribute vectors and low-level features.

7. The zero-sample performance degradation evaluation method for intelligent decision-making in compressor water washing according to claim 6, characterized in that, The formula for calculating the generative adversarial loss function is as follows: in, D represents the generative adversarial loss function. b (·) represents the discrimination result of the high-level features output by the adversarial discriminator. This represents the high-level features output by the trained feature extractor. This represents the high-level characteristics generated by the converter output, where λ represents the penalty coefficient. α~U(0,1) represents the weighting coefficients, U(0,1) represents a uniform distribution defined on the interval [0,1], and ∥·∥2 represents the L2 norm. express about The gradient of , E represents the mathematical expectation, and represents the average loss over all the aforementioned sample pairs; The formula for calculating the second label discrimination loss function is as follows: Among them, L A ′ represents the second-label discriminant loss function. Indicates high-level samples Running data Corresponding attribute vector The actual value of the j-th attribute.

8. The zero-sample performance degradation evaluation method for intelligent decision-making in compressor water washing according to claim 1, characterized in that, Step (4) specifically includes: First, training set data x is obtained by sampling from the training set. tr and its corresponding degradation level label y tr and attribute vector a s Randomly select attribute vector a from the undegraded levels. u and its corresponding degradation level label y u Then, the trained feature extractor is used to extract the training set data x. tr The domain-invariant known degradation level features 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 differences between the two sets of features are calculated, and then the attribute differences are concatenated with random noise and input into the trained generator to obtain the generated feature differences. The generated feature differences are then added to the training set features and input into the trained converter to obtain the generated unseen degradation level features. Finally, the generated unseen degradation level features and the training set features are input into the classifier to obtain the corresponding predicted degradation level labels. Using the cross-entropy loss function, based on 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. tr Calculate the supervised training loss function of the classifier; adjust the parameters of the classifier with the goal of minimizing the supervised training loss function of the classifier to obtain the trained classifier; the trained classifier can simultaneously classify features with known degradation levels and features without known degradation levels.

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

Images