Heavy-duty gas turbine compressor scale deposition detection method based on liquid neural network

Abstract

The invention discloses a heavy duty gas turbine compressor scale deposition detection method based on a liquid neural network, and relates to the technical field of anomaly detection, and the method comprises the following steps: collecting operation data for correlation analysis; the selected feature data is preprocessed; extracting time characteristics of the operation parameters of the gas compressor; calculating weights of interaction among the parameters, and outputting spatial correlation; constructing a space-time fusion model, and inputting normal operation data to train the model; and a parameter prediction sequence is output, a multi-dimensional residual vector of a predicted value and an actual value is calculated, and by constructing a dynamic threshold function, when the model converges and residual distribution deviates from a normal working condition confidence interval, a gas compressor fouling anomaly detection result is output. According to the invention, through the dynamic topological structure of the liquid neural network and in combination with a spatial-temporal characteristic collaborative mining mechanism, quantitative evaluation of the scale deposition degree of the gas compressor is realized, and the precision and robustness of gas compressor scale deposition anomaly detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal detection of heavy-duty gas turbine compressor systems, and particularly to a method for detecting fouling of a heavy-duty gas turbine compressor based on a liquid neural network. Background Art

[0002] As a core power equipment in the fields of aviation and power generation, the performance and operation safety of a heavy-duty gas turbine directly affect the system reliability. The performance degradation of a gas turbine is mainly due to the performance decline of the compressor and the turbine, especially the problem of compressor fouling. Fouling is caused by the accumulation of impurities in the air on the surfaces of the compressor blades and flow passages, resulting in changes in the air flow geometry and surface quality, and thus affecting the performance of the compressor and the entire engine. The main maintenance strategies for compressor fouling include on-line / off-line cleaning. Optimizing the cleaning cycle is crucial for the economy and safety of gas turbine operation because it requires a trade-off between performance improvement and corresponding costs. Therefore, how to accurately predict the degradation trend of the compressor and adopt an efficient and economical water washing strategy has become a technical bottleneck that urgently needs to be broken through for the safe and economic operation of current gas turbines.

[0003] Compressors often have difficulties in mechanism analysis and strong dynamics due to reasons such as changes in operating conditions, multi-variable coupling, and external environmental interference, resulting in difficulties in accurate modeling. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for detecting fouling of a heavy-duty gas turbine compressor based on a liquid neural network, to solve the problems proposed in the above background art, effectively simulate the dynamics of compressor operation, extract spatial features using a graph attention mechanism, and adaptively track the drift of the residual distribution during the compressor fouling process through a dynamic threshold algorithm, so as to improve the prediction accuracy of compressor fouling anomalies.

[0005] To achieve the above purpose, the present invention provides a method for detecting fouling of a heavy-duty gas turbine compressor based on a liquid neural network, including the following steps: Step S1: Due to problems such as multi-variables and strong coupling in the compressor, the Spearman correlation coefficient method is used to perform a correlation analysis on the compressor operation data, and variables with strong correlations are selected for compressor modeling; Step S2: Preprocess the selected feature data and divide the data into a training set and a test set; Step S3: Due to the difficulties in mechanism analysis and strong dynamics of the compressor, it is difficult to accurately model. Therefore, a liquid neural network with good modeling ability for complex dynamic data is used as a time model to extract the time features of the compressor operation parameters; Step S4: Considering the coupling phenomenon in the compressor operation parameters, calculate the weights of the interactions between the compressor operation parameters, and use the graph attention mechanism to output spatial correlations; Step S5: Construct a spatio-temporal fusion model for training, calculate the multi-dimensional residual vector, and construct a dynamic threshold function; Step S6: Observe whether the model tends to converge and whether the residual distribution deviates from the confidence interval of the normal operating condition. If the model converges and the residual distribution deviates from the confidence interval of the normal operating condition, output the abnormal detection result of the compressor fouling. Otherwise, repeat steps S3 - S5.

[0006] In the present invention, by using a liquid neural network for modeling, the hierarchical depth features of a static neural network and the temporal dynamic characteristics of a recurrent neural network (RNN) are converted into a continuous vector field, thereby realizing parameter sharing, adaptive calculation of non-uniformly sampled data, and function approximation. Therefore, it has important theoretical value and practical significance to use the liquid neural network method to study the abnormal detection of the heavy-duty gas turbine compressor system.

[0007] Preferably, in step S1, the Spearman correlation coefficient method is used to perform a correlation analysis on the compressor operation data, and variables with strong correlations are selected for compressor modeling. The specific formula is: ; where, is the total number of observation objects, is the rank difference of the i th observation object in two variables. If represents a perfect positive correlation, represents a perfect negative correlation, represents no monotonic correlation relationship.

[0008] Preferably, the preprocessing of the operation data collected in step S2 specifically includes collecting the operation data under normal and abnormal states at different measuring points in the heavy-duty gas turbine compressor system, and preprocessing the missing values and abnormal values of the data.

[0009] Preferably, in step S3, a liquid neural network is used as a time model to extract time features. The core lies in using a dynamically changing time constant to control the state change rate of each neuron. This structure enables each neuron to adaptively change at different time points according to the input features. The dynamic equation formula is as follows: ; where, t is the time step, is the hidden state of the liquid time constant layer of the neuron; is the liquid time constant parameter vector, which controls the rate of neuron state update; is a neural network with as the parameter; is the exogenous input of the neural network; A is a bias vector.

[0010] The time constant parameter not only determines the change rate of the hidden state but is also closely related to the input features. Specifically, the liquid time constant can be expressed by the following formula: ; where, is the time-varying factor; is a neural network with as the parameter; is the exogenous input of the neural network.

[0011] The approximate closed-form solution of the liquid neural network can be expressed by the following formula: ; is shown as: ; where, B is the parameter vector of the decomposition; is the liquid time constant parameter vector; is a neural network with as the parameter; is the exogenous input of the neural network; A is a bias vector; is the Hadamard product.

[0012] Preferably, the specific steps of step S4 are as follows: Step S41: Calculate the attention value between the compressor operation parameters; Step S42: Normalize the adjacent attention information of each node; Step S43: Adopt the multi-head attention mechanism for feature aggregation; Step S44: Output the spatial correlation through the graph attention mechanism.

[0013] Preferably, the formula for calculating the attention value in step S41 is as follows: ; where, a represents the attention network; respectively represent nodes i and j ; W represents the weight sharing matrix.

[0014] Preferably, the formula for normalization processing in step S42 is as follows: ; Among them, represents the attention coefficient between nodes i and j ; represents the attention value between nodes i and j ; represents the attention value between nodes i and k ; represents the neighbor nodes of node i .

[0015] Preferably, in step S43, a multi-head attention mechanism is used for aggregation, and the formula is as follows: ; Among them, represents the feature representation of node i ; represents concatenating the outputs of heads; represents an arbitrary activation function; represents the attention coefficient of the th head; represents the weight matrix of the th head; represents the feature representation of node j .

[0016] Preferably, the graph attention formula in step S44 is as follows: ; Among them, represents the attention coefficient between nodes i and j ; represents the neighbor nodes of node i ; represents the activation function; represents a single-layer feedforward neural network, T represents the transpose; W represents the shared weight matrix; represents the feature representation of node i ; represents the feature representation of node j ; represents the feature representation of node k .

[0017] Preferably, the specific steps of step S5 are as follows: Step S51: Input the data into the spatio-temporal fusion model for training by using a sliding window mechanism; Step S52: Calculate the multi-dimensional residual vector and construct a dynamic threshold function based on the historical normal operating condition residual samples; Step S53: Set the trigger condition for outputting the detection result of compressor fouling abnormality.

[0018] Preferably, the formula of the sliding window function in the step S51 is as follows: ; where, represents the window length; represents the selected monitoring parameter dimension; represents t the observation vector at time

[0019] Input the current window data into the spatio-temporal fusion model , and output the prediction sequence. The formula is as follows: ; where, represents the prediction length.

[0020] Preferably, the formula for calculating the multi-dimensional residual vector in the step S52 is as follows: ; where, represents the covariance matrix of the normal data residuals in the training stage; represents the Mahalanobis distance metric.

[0021] The formula for constructing the probability distribution model by kernel density estimation is as follows: ; where, the kernel function selects the Gaussian kernel. The formula is as follows: ; The bandwidth matrix H is determined by the Silverman criterion. The formula is as follows: ; where, represents the standard deviation estimate of the i th dimension of the residuals; represents the standard deviation estimate of the j th dimension of the residuals; Q represents the number of residual samples.

[0022] Set the confidence level , and calculate the dynamic threshold. The formula is as follows: ; where, represents the inverse function of the cumulative distribution function.

[0023] Preferably, the trigger condition formula for outputting the fouling abnormality detection result of the compressor in step S53 is: ; Output the abnormal detection result of compressor fouling and give the abnormal confidence level: ; Wherein, represents the sensitivity adjustment factor (default value 0.5).

[0024] Therefore, the present invention adopts the above-mentioned compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks. Through the dynamic topological structure of the liquid neural network and the collaborative mining mechanism of spatio-temporal features, the quantitative evaluation of the compressor fouling degree is realized, and the accuracy and robustness of the abnormal detection of compressor fouling are improved.

[0025] Next, through the drawings and embodiments, the technical solution of the present invention will be further described in detail. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flowchart of an embodiment of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention; Figure 2 is a Spearman correlation coefficient heat map of an embodiment of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention; Figure 3 is a schematic diagram of a liquid neural network of an embodiment of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention; Figure 4 is a schematic diagram of a graph attention network of an embodiment of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention; Figure 5 is a prediction curve graph of the liquid neural network of an embodiment of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention for the output parameters of the compressor. Among them, (a) is the prediction curve graph of the compressor exhaust pressure; (b) is the prediction curve graph of the compressor exhaust temperature; Figure 6 is a histogram of residual distribution and a kernel density curve graph of an embodiment of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention; Figure 7 is a comparison graph of the residual value distributions of the compressor fouling detection method for heavy-duty gas turbines based on liquid neural networks of the present invention and other detection methods; Figure 8This is a schematic structural diagram of a heavy-duty gas turbine for an embodiment of the fouling detection method for the compressor of a heavy-duty gas turbine based on a liquid neural network. Detailed implementation manners

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those of ordinary skill in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0029] Embodiment Please refer to Figure 1 , the present invention provides a fouling detection method for the compressor of a heavy-duty gas turbine based on a liquid neural network, including the following steps: Step S1: Use the Spearman correlation coefficient method to perform a correlation analysis on the compressor operation data, and select variables with strong correlations for compressor modeling. The specific variables are shown in Table 1 below; Table 1 Compressor operation related variables of a heavy-duty gas turbine

[0030] Please refer to Figure 2 , in step S1, use the Spearman correlation coefficient method to perform a correlation analysis on the compressor operation data selected in Table 1, and select variables with strong correlations for compressor modeling. The specific formula is: ; Wherein, is the total number of observation objects, is the rank difference of the i th observation object in the two variables. If represents a perfect positive correlation, represents a perfect negative correlation, represents no monotonic correlation relationship.

[0031] Figure 2The results show that there is a strong coupling relationship among the selected operating parameters under the same working conditions, which needs to be paid special attention to in subsequent modeling or analysis.

[0032] Step S2: Collect operation data for preprocessing, specifically including collecting the operation data of different measuring points in the compressor system of a heavy-duty gas turbine under normal and abnormal states, and preprocessing the missing values and abnormal values of the data. After preprocessing, the vibration data is divided into a training set and a test set.

[0033] Step S3: Use a liquid neural network as a time model to extract the time features of the compressor operation parameters; Step S3 uses Figure 3 The shown liquid neural network as a time model to extract time features. Its core lies in using a dynamically changing time constant to control the state change rate of each neuron. This structure enables each neuron to adaptively change at different time points according to the input features. Its dynamic equation formula is as follows: ; Among them, t is the time step, is the hidden state of the liquid time constant layer of the neuron; is the liquid time constant parameter vector, which controls the rate of neuron state update; is a neural network with as the parameter; is the exogenous input of the neural network; A is a bias vector.

[0034] The liquid time constant parameter not only determines the change rate of the hidden state but also is closely related to the input features. Specifically, the liquid time constant can be expressed by the following formula: ; Among them, is the time-varying factor; is a neural network with as the parameter; is the exogenous input of the neural network, t is the time step.

[0035] The approximate closed-form solution of the liquid neural network can be expressed by the following formula: ; The display expression is: ; Among them, B is the parameter vector of the decomposition; is the parameter vector of the liquid time constant; is a neural network with as the parameter; is the exogenous input of the neural network; A is a bias vector; is the Hadamard product.

[0036] Step S4: Calculate the attention weights of the interactions between the compressor operating parameters, perform normalization processing, then use the multi-head attention mechanism for feature aggregation, and finally output the spatial correlation, which significantly improves the accuracy and robustness of the compressor fouling anomaly detection. The specific steps are as follows: Step S41: Calculate the attention values between the compressor operating parameters. The formula is as follows: ; where, a represents the attention network; respectively represent the nodes i and j ; W represents the weight sharing matrix.

[0037] Step S42: Perform normalization processing on the adjacent attention information of each node. The formula is as follows: ; where, represents the attention coefficient between the nodes i and j ; represents the attention value between the nodes i and j ; represents the attention value between the nodes i and k ; represents the neighbor nodes of the node i .

[0038] Step S43: Use the multi-head attention mechanism for feature aggregation. The formula is as follows: ; where, represents the feature representation of the node i ; represents concatenating the outputs of heads; represents an arbitrary activation function; represents the th head's attention coefficient; represents the th head's weight matrix; represents the node jFeature representation.

[0039] Step S44: Output spatial correlation through the graph attention mechanism, and the formula is as follows: ; Among them, represents the attention coefficient between nodes i and j ; represents the neighbor nodes of node i ; represents the activation function; represents a single-layer feedforward neural network, T represents the transpose; W represents the shared weight matrix; represents the feature representation of node i ; represents the feature representation of node j ; represents the feature representation of node k ;

[0040] The schematic diagram of the graph attention network is as shown in Figure 4 . By calculating the attention weights between each node and its neighbor nodes and using the multi-head attention mechanism to aggregate the neighbor node features, dynamic feature weighted aggregation based on the attention coefficient is achieved.

[0041] Step S5: Construct a spatio-temporal fusion model for training, calculate the multi-dimensional residual vector, and adaptively track the drift of the residual distribution during the fouling process of the compressor through the dynamic threshold algorithm. The specific steps are as follows: Step S51: Input the data into the spatio-temporal fusion model for training using the sliding window mechanism. The sliding window function is defined as follows: ; Among them, represents the window length; represents the selected monitoring parameter dimension; represents t the observation vector at time

[0042] Input the current window data into the spatio-temporal fusion model , and generate the prediction sequence as shown in Figure 5 : ; Among them, represents the prediction length.

[0043] Figure 5The results show that the model integrating the liquid neural network and the graph attention mechanism achieves accurate prediction of the compressor exhaust pressure and exhaust temperature, laying a model foundation for the subsequent quantitative evaluation of the compressor fouling degree.

[0044] Step S52: Refer to Figure 6 , calculate the multi-dimensional residual vector, and the formula is as follows; ; Among them, represents the covariance matrix of the normal data residuals in the training stage; represents the Mahalanobis distance metric.

[0045] Based on the historical normal operating condition residual samples , the formula for constructing the probability distribution model using kernel density estimation is as follows: ; Among them, the kernel function selects the Gaussian kernel, and the formula is as follows: ; The bandwidth matrix H is determined by the Silverman criterion, and the formula is as follows: ; Among them, represents the standard deviation estimate of the i -dimensional residual; represents the standard deviation estimate of the j -dimensional residual; Q represents the number of residual samples.

[0046] Figure 6 The results show that most of the residuals are concentrated near zero and show relatively obvious peaks, indicating that the predicted values are close to the true values. The residual probability density distribution constructed using kernel density estimation is approximately a normal distribution, indicating that the prediction model is generally accurate.

[0047] Set the confidence level , calculate the dynamic threshold, and the formula is as follows: ; Step S53: When m consecutive windows satisfy , output the abnormal detection result of the compressor fouling and give the abnormal confidence level, and the formula is as follows: ; Among them, represents the sensitivity adjustment factor (default value 0.5).

[0048] Step S6: Observe whether the model tends to converge and whether the residual distribution deviates from the confidence interval of the normal operating condition. If the model converges and the residual distribution deviates from the confidence interval of the normal operating condition, output the abnormal detection result of the compressor fouling; otherwise, repeat steps S3 - S5; See Figure 7 , which is a comparison of the residual distribution of the model of the present invention and two improved LSTM models. Figure 7 The results show that the error of the BiLSTM model is large and relatively dispersed, while the errors of the xLSTM and the model of the present invention are small and concentrated. Compared with the xLSTM model, the median of the model of the present invention is closer to 0 and the error range is smaller. Therefore, compared with other detection methods, the diagnostic accuracy of the present invention is higher.

[0049] Figure 8 The figure shows a schematic diagram of the structure of a certain type of single - shaft heavy - duty gas turbine. Compressor fouling is the main reason for the performance degradation of heavy - duty gas turbines. Studying its abnormal detection technology is of great significance.

[0050] It can be seen from this embodiment that this method realizes the dynamic update of feature information through a liquid - state neural network, strengthens the robustness and generalization ability of the model, uses a graph attention mechanism to output the spatial correlation between operating parameters, and constructs a dynamic threshold based on a multi - dimensional residual vector, improving the diagnostic accuracy of the model.

[0051] Therefore, the present invention adopts the above - mentioned method for detecting compressor fouling of heavy - duty gas turbines based on a liquid - state neural network. Through the dynamic topological structure of the liquid - state neural network and the combination of a spatio - temporal feature collaborative mining mechanism, it realizes the quantitative evaluation of the degree of compressor fouling and improves the accuracy and robustness of abnormal detection of compressor fouling.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A heavy-duty gas turbine compressor fouling detection method based on liquid neural network, characterized in that: The following steps are involved: Step S1: using the Spearman correlation coefficient method to perform correlation analysis on the compressor operating parameters to obtain characteristic data, and selecting characteristic data with strong correlation for compressor modeling; Step S2: pre-processing the highly correlated feature data and dividing them into a training set and a test set; Step S3: using a liquid neural network as a time model to extract the time characteristics of the compressor operating parameters; Step S4: Calculate the weights of the interactions between the extracted compressor operating parameters, and output spatial correlation using a graph attention mechanism; Step S5: construct a spatiotemporal fusion model for training, calculate a multidimensional residual vector, and construct a dynamic threshold function; Step S6: Observe whether the space-time fusion model tends to converge and whether the residual distribution deviates from the normal operating condition confidence interval. If the space-time fusion model converges and the residual distribution deviates from the normal operating condition confidence interval, the compressor fouling anomaly detection result is output, otherwise repeat steps S3-S5.

2. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 1 is characterized in that: The method of using the Spearman correlation coefficient method to perform correlation analysis on the compressor operating parameters in step S1 is calculated according to the following formula: in, is the total number of observations, For the i The rank difference of an observation in two variables is indicates a completely positive correlation. indicates a completely negative correlation. Indicates that there is no monotonic correlation.

3. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 2 is characterized in that: In step S3, the liquid neural network is used as a time model to extract the time characteristics of the compressor operating parameters by using a dynamically changing time constant to control the state change rate of each neuron. This structure enables each neuron to change adaptively at different time points according to the input characteristics. The specific formula is as follows: in, t is the time step, is the hidden state of the liquid time constant layer of the neuron; is the liquid time constant parameter vector, which controls the rate at which the neuron state is updated; It is a A neural network with parameters is the exogenous input to the neural network; A is a bias vector; Liquid time constant It can be expressed by the following formula: in, is a time-varying factor; It is a A neural network with parameters is the exogenous input to the neural network; The approximate closed-form solution of the liquid neural network can be expressed by the following formula: The display expression is: in, B yes Decomposed parameter vector; is the liquid time constant parameter vector; It is a A neural network with parameters is the exogenous input to the neural network; A is a bias vector; is the Hadamard product.

4. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 3 is characterized in that: The step S4 comprises: Step S41: calculating the attention value between compressor operating parameters; Step S42: normalizing the adjacent attention information of each node; Step S43: adopting a multi-head attention mechanism to perform feature aggregation; Step S44: Output spatial correlation through the graph attention mechanism.

5. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 4 is characterized in that: The formula for outputting spatial correlation through the graph attention mechanism in step S44 is as follows: in, Representation Node i and j The attention coefficient between Representation Node i Neighbor nodes of represents the activation function; represents a single-layer feedforward neural network, T represents transpose; W represents the shared weight matrix; Representation Node i The feature representation of Representation Node j The feature representation of Representation Node k The feature representation of .

6. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 5 is characterized in that: The step S5 comprises: Step S51: inputting the data into the spatiotemporal fusion model for training using a sliding window mechanism; Step S52: Calculate a multi-dimensional residual vector, and construct a dynamic threshold function based on historical normal operating condition residual samples; Step S53: setting a trigger condition for outputting the compressor fouling abnormality detection result.

7. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 6 is characterized in that: The function definition of the sliding window mechanism in step S51 is as follows: in, Indicates the window length; Indicates the selected monitoring parameter dimension; express t The observation vector at time instant; Input the current window data into the spatiotemporal fusion model , generate the prediction sequence: in, Indicates the predicted length.

8. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 7 is characterized in that: The expression formula for calculating the multidimensional residual vector in step S52 is: in, Represents the covariance matrix of the residuals of normal data during the training phase; represents the Mahalanobis distance metric; Based on historical normal operating condition residual samples The formula for constructing the dynamic threshold function is as follows: Among them, the kernel function Select the Gaussian kernel, the formula is as follows: Bandwidth Matrix H Determined by Silverman's criterion: in, Indicates i The standard deviation estimate of the dimensional residual; Indicates j The standard deviation estimate of the dimensional residual; Q represents the number of residual samples; Setting the confidence level , calculate the dynamic threshold, the formula is as follows: in, Represents the inverse of the cumulative distribution function.

9. The heavy-duty gas turbine compressor fouling detection method based on liquid neural network according to claim 8, characterized in that: The triggering condition for outputting the compressor fouling abnormality detection result in step S53 is: m The window satisfies the following formula: Output the abnormal detection results of compressor fouling and give the abnormal confidence level to represent the confidence interval: in, Represents the sensitivity adjustment factor.

Images