A knowledge-based hybrid drive-based operation status monitoring method for gas turbines

Abstract

This invention discloses a knowledge-condition hybrid-driven operational status monitoring method for gas turbines. The method constructs a knowledge matrix for gas turbines by combining expert knowledge with correlation analysis between variables, and divides the overall object into multiple sub-modules, using the knowledge matrix to guide the selection of variables for each sub-module. Based on this, a condition-driven approach is introduced, decomposing a large-scale non-stationary continuous process into several conditional modes, and establishing a state monitoring model for each conditional mode to achieve operational status monitoring under varying operating conditions. Finally, the weight coefficients of the sub-modules are determined by combining conditional indicator variables, and the monitoring results of the sub-modules are integrated to achieve comprehensive monitoring of the operational status of complex industrial equipment. This invention not only allows expert knowledge to be better applied to the monitoring of non-stationary processes, but also enhances the accuracy and reliability of actual online monitoring of gas turbines, providing an efficient and reliable new approach for the operational status monitoring of complex industrial equipment.

Description

Technical Field

[0001] This invention belongs to the field of monitoring the operating status of complex industrial equipment operating on a large scale and with non-stationary conditions. In particular, it relates to a method that combines knowledge-driven and condition-driven approaches, constructs a knowledge matrix through expert knowledge and variable correlation analysis, divides the overall object into multiple sub-components and establishes condition-driven monitoring models for each, and finally fuses the monitoring results based on the values ​​of condition indicator variables to achieve intelligent monitoring of the operating status of a gas turbine. Background Technology

[0002] A gas turbine is an internal combustion engine that generates useful work by rotating turbine blades through fuel combustion. It is the most efficient type of heat-to-work conversion power generation equipment to date, widely used in aviation, shipbuilding, and power generation, playing a crucial role as core equipment. In practical applications, taking power generation as an example, gas turbine generator sets often participate in grid peak shaving, leading to variable loads and frequent start-ups and shutdowns. Their normal operation includes multiple steady-state and transient conditions, exhibiting significant variable operating condition characteristics. Distinguishing between fault operation and normal operation under these variable conditions is a major challenge in gas turbine condition monitoring. Furthermore, as a complex industrial system, the gas turbine contains multiple sub-components and numerous measuring points. Interactions between components and measuring points during operation create coupling relationships, significantly increasing the difficulty of condition monitoring. Currently, research on gas turbine monitoring technology in China is insufficient, and the gas turbine industry is still immature. Research on gas turbine monitoring technology urgently needs to be advanced.

[0003] Existing gas turbine operating condition monitoring methods are mainly divided into digital signal analysis methods and data-driven methods. Early condition monitoring methods used digital signal analysis, employing methods such as autocorrelation coefficients and Fourier transforms to analyze time-domain and frequency-domain signals of the gas turbine, including pressure, vibration, and sound. However, digital signal analysis methods have high requirements for sensor data and require the acquisition of compressor operating data under fault conditions, which can cause some damage to the equipment. With the rapid development of machine learning, data-driven monitoring methods have gradually become mainstream. These methods build and update monitoring models based on operating data, and the types of models are very diverse, such as Bayesian models, Support Vector Machines (SVM), Extreme Learning Machines (ELM), Stacked Denoising Autoencoders (SDAE), and fuzzy clustering algorithms. However, current data-driven monitoring methods often assume stable operating conditions for the target object, meaning they only monitor a specific operating condition. Since the actual operating conditions of gas turbines are often variable and include multiple transition periods, exhibiting typical large-scale non-stationary characteristics, these methods are not suitable for monitoring the operating condition of gas turbines in large-scale non-stationary scenarios, and therefore cannot be well applied to actual industrial processes.

[0004] In recent years, condition-driven operational status monitoring methods have received increasing attention to address the challenges of monitoring operational status in large-scale non-stationary scenarios. These methods divide a large-scale non-stationary transient continuous process into several conditional modes and conduct detailed process analysis for each mode to comprehensively describe the potential distribution of the data. This allows the model to adapt to a wide range of variable inputs and demonstrates good monitoring performance under various operating conditions. While some research has been conducted in the industrial field on the state monitoring of large-scale non-stationary processes—for example, Zhao et al. reconstructed a non-stationary sequence in the time dimension into a stationary sequence in the conditional dimension and defined a new Bayesian inference statistic for online monitoring after analyzing the static and dynamic information of the process—these methods, despite some successful industrial applications, present significant challenges for large, complex industrial equipment like gas turbines with high complexity and strong coupling. The obvious coupling between components and the mutual influence between measured variables mean that directly using all variables to build a monitoring model on the entire object cannot accurately reflect the operational status of sub-components, resulting in inaccurate and uninterpretable monitoring results. In such cases, purely condition-driven monitoring methods may not achieve satisfactory operational status monitoring results.

[0005] In summary, real-world complex industrial processes often possess intricate internal structures, consisting of multiple sub-components with complex coupling relationships between their variables. Furthermore, their operational states are typically transient, remaining in a particular state for only a short period and potentially changing continuously, exhibiting typical dynamic characteristics. Therefore, a combined knowledge-driven and condition-driven operational state monitoring method should be designed. This method leverages expert knowledge to divide processes into sub-modules and constructs a knowledge matrix to guide the establishment of conditional monitoring models. This reduces the impact of inter-component coupling and widespread non-stationary characteristics on monitoring results, thereby improving the accuracy and reliability of the monitoring outcomes. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing operational status monitoring technologies for large-scale non-stationary processes by providing a hybrid operational status monitoring method combining knowledge matrix and condition-driven approaches. This method constructs a knowledge matrix by integrating expert knowledge and the correlations between measured variables, and then divides the overall process into sub-modules. The main variables and their related variables for each sub-module are selected based on the knowledge matrix. Furthermore, this method introduces a data reconstruction strategy to reconstruct the data of the large-scale non-stationary process into a stationary sequence in the conditional dimension, extracting static and dynamic information. A condition-driven monitoring model is then established for each sub-module. By comparing the correlation coefficients between the main variables of each sub-module and the main variables of the overall process, the method determines the weight of the monitoring results for each sub-module, thus deriving the final monitoring results. This invention provides a novel analytical perspective for operational status monitoring of large-scale non-stationary processes. It not only incorporates mechanistic knowledge of gas turbines in the real world but also enhances the reliability of monitoring results by combining knowledge-driven and condition-driven approaches. This helps industrial engineers accurately judge the operational status of gas turbines during large-scale non-stationary transient continuous processes, promptly detect faults, and thus ensure the safety of actual production.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A knowledge-based hybrid-driven operation status monitoring method for gas turbines, specifically including:

[0009] Building a monitoring model includes the following steps:

[0010] (1) Obtain N samples of the normal continuous operation process of the gas turbine system, each sample containing J measured variables;

[0011] (2) Calculate the vector X formed by N samples of any pairwise measured variables. m X n The correlation coefficient between them, corr(X) m X n Based on the calculated correlation coefficients and combined with the mechanistic knowledge of the object, a knowledge matrix M is constructed, where the element in the m-th row and n-th column of the knowledge matrix is ​​M. mn Representative variable X m and X n Whether there is a relationship; 1 indicates a relationship, 0 indicates no relationship;

[0012] (3) Based on the mechanistic knowledge of the gas turbine and the physical structure of the object, the whole object is divided into S sub-modules; and based on the constructed knowledge matrix M and the object's mechanistic knowledge, a condition indicator variable X that can best indicate the process operating state is selected for each sub-module s.s，c and other related variables X associated with this variable. s，o Together they form a two-dimensional sample matrix X of submodule s. s =[X s，c X s，o Simultaneously, based on the constructed knowledge matrix M and the object mechanism knowledge, a conditional indicator variable X is selected that can best indicate the overall process operation state of the object. c ;

[0013] (4) For the two-dimensional data matrix X of each sub-module s Calculate the first difference between each sample and the previous sample. Each sample and its corresponding first-order difference combination are used as new samples. All new samples result in a dynamic two-dimensional sample matrix.

[0014] (5) For each submodule sample matrix X s , The samples are rearranged in ascending order of condition indicator variable values ​​and divided into K groups, forming K static condition slices. and K dynamic condition slices Where N k，s Let k be the number of samples in the k-th condition slice of submodule s, where k = 1, ..., K; then, mean-removal processing is performed on each condition slice to obtain the data matrix of standardized condition slices. and The standardized static condition pieces are merged to obtain C. s There are several static condition segments, each consisting of multiple consecutive standardized static condition pieces. The control limit calculated based on the PCA model for each static condition segment should be less than the product of the control limit calculated based on the PCA model for each static condition piece within the segment and the constant scaling factor γ. A monitoring model and control limit are established for each static condition segment. c，s ;

[0015] C is also divided based on the breakpoints of the static condition segment. s Each dynamic condition segment is defined, and a monitoring model and control limits are established for each segment.

[0016] (6) Divide the overall conditional indicator variable into Z intervals, and calculate the conditional indicator variable X for each sub-module within each interval z. s，c With the overall condition indicator variable X c The weight of the correlation coefficient serves as the weighting coefficient for each submodule when the overall conditional indicator variable is in the interval z.

[0017] Online monitoring includes the following steps:

[0018] Real-time acquisition of the test sample x new (1×J), and processed to obtain the first-order difference. and its samples x in each submodule new，s Determine the interval z to which the sample belongs based on the overall conditional indicator variable value. new Based on the submodule condition indicator variable values ​​of the sample, it is determined to belong to the static and dynamic condition segments, and the monitoring model, control limits, and corresponding interval z of the static and dynamic condition segments are used as the basis for these determinations. new The weighting coefficients are used to calculate the monitoring indicators and control limits of the current sample under test. The gas turbine is considered to be operating under normal conditions if and only if both static and dynamic monitoring indicators are less than their control limits; otherwise, the gas turbine is considered to have malfunctioned.

[0019] Furthermore, the measured variables include multiple of the following: pre-module natural gas volume flow rate, pre-module natural gas mass flow rate, compressor inlet temperature, compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust bearing generator end temperature, compressor thrust bearing gas turbine end temperature, compressor bearing vibration, compressor side shaft vibration, compressor inlet pressure differential, compressor anti-icing device inlet electric regulating valve position, gas turbine exhaust average temperature, gas turbine side shaft vibration, combustion chamber pressure differential, gas turbine cooling air regulating valve position 1, gas turbine cooling air regulating valve position 2, gas turbine humming, gas turbine speed, gas turbine power, gas turbine second-stage stationary vane ring chamber cooling air pressure, and gas turbine third-stage stationary vane ring chamber cooling air pressure.

[0020] Furthermore, the division of the overall object into S sub-modules specifically refers to:

[0021] The overall object is divided into three sub-modules: compressor, combustion chamber, and gas turbine.

[0022] Furthermore, the establishment of a monitoring model and control limits for each static condition segment (Ctrl) c，s Specifically as follows:

[0023] Establish a PCA model for each static condition segment. in, It is the data matrix of the c-th static condition segment in the s-th module. express The corresponding principal components, This represents the PCA transformation matrix corresponding to the c-th static condition segment of the s-th module;

[0024] For each static condition segment, a Gaussian Mixture Model (GMM) is established in the feature space, and the FJ algorithm is used to determine the number of Gaussian elements v. cThen, the EM algorithm is used to estimate the GMM parameters. Let be the GMM parameter of the i-th Gaussian element of the static condition segment c; This refers to the monitoring model for static condition segment c;

[0025] Constructing the probability density function g based on the PCA model with static condition segments and GMM parameters:

[0026]

[0027] in, Let be the parameter of the probability density function, and be the number of Gaussian elements in the feature space of the static condition segment c. Let be the prior probability of the i-th Gaussian element in condition segment c. Let be the GMM parameter of the i-th Gaussian element of condition segment c;

[0028] Calculate the Bayesian inference distance index for static condition segment c:

[0029]

[0030]

[0031] Where, N c，s Let be the number of samples in condition segment c within submodule s. Let be the posterior probability that sample j belongs to the i-th Gaussian component. For T j，c，s Mahalanobis distance to the i-th Gaussian component; T j，c，s express The j-th row; based on the confidence level 1-α, the BID is calculated using the kernel density estimation method. c，s Control limits Ctrl c，s .

[0032] Furthermore, the step of determining the static and dynamic condition segments to which the sample belongs based on the sub-module condition indicator variable values, and then basing this determination on the monitoring model, control limits, and corresponding interval z of the static and dynamic condition segments. new The weighting coefficients are used to calculate the monitoring indicators and control limits of the current sample to be tested, as follows:

[0033] Calculate the monitoring statistics of the sample to be tested in the submodule:

[0034]

[0035]

[0036] Among them, c sIt is the condition segment number of the sample to be tested in submodule s. c s The monitoring model corresponding to the condition segment The feature values ​​of the sample to be tested are specifically:

[0037] According to the interval z new The submodule weight coefficients are used to calculate the static monitoring index BID. new Dynamic monitoring indicators and the corresponding control limit Ctrl new ,

[0038]

[0039]

[0040]

[0041]

[0042] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a new research approach for the state monitoring of large-scale non-stationary continuous processes. By designing a knowledge-driven condition indicator variable selection strategy, the concept of knowledge-driven approaches is introduced into condition-driven monitoring methods for the first time. A knowledge matrix is ​​constructed by combining object expert knowledge and the correlation between variables, and sub-modules are divided. Under the guidance of the knowledge matrix, the condition variables and related variables of each sub-module are determined, thus providing a foundation for further analysis. By introducing a data reconstruction strategy, the data is rearranged along the condition direction, dividing it into multiple condition segments, and a monitoring model is further established on each condition segment, making the algorithm applicable to various working conditions. Considering the differences in the impact of sub-modules on the overall operating state under different working conditions, an intelligent fusion method for sub-module monitoring results is designed. The correlation between condition variables under different working conditions is used to determine the weight coefficients of sub-modules to fuse the overall monitoring results, realizing intelligent monitoring of the operating state through the fusion of knowledge matrix and condition-driven approaches. The proposed method has been successfully applied in detailed experimental studies in actual industrial processes. This method enhances the application of prior knowledge of specific objects by establishing a knowledge matrix for a wide range of non-stationary sequences. By combining knowledge-driven and condition-driven approaches, it avoids the coupling problem between components in traditional monitoring methods, improves the accuracy and reliability of process monitoring, and can ultimately be applied to gas turbine combined cycle power generation sites to ensure the safe and reliable operation of gas turbines. Attached Figure Description

[0043] Figure 1 This is a flowchart of the operation status monitoring method that combines knowledge matrix and condition-driven approaches according to the present invention;

[0044] Figure 2 This is a variable list and knowledge matrix diagram in a specific embodiment of the present invention, wherein (a) is the name and corresponding number of the variable being tested, and (b) is the knowledge matrix;

[0045] Figure 3 These are the indications of the condition indicators for each component;

[0046] Figure 4 These are the monitoring results of the method of the present invention in specific embodiments of the present invention and the comparison results with traditional monitoring methods, wherein (a) is the monitoring result of static indicators of fault cases, (b) is the monitoring result of dynamic indicators of fault cases, and (c) is the comparison result of the method proposed in the present invention with traditional monitoring methods. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and specific examples.

[0048] Combined cycle gas turbine generator sets represent a highly complex industrial process characterized by time-varying, dynamic, and non-stationary characteristics. Heavy-duty gas turbines for power generation consist of three core components: a compressor, a combustion chamber, and a gas turbine. The compressor and turbine are typically multi-stage axial-flow designs. The working principle is as follows: air enters through the gas turbine inlet, where the high-speed rotation of the multi-stage rotor blades in the compressor accelerates and pressurizes the airflow, which is then sent to the combustion chamber. Fuel (gaseous or liquid fuel) is then injected into the combustion chamber and mixed with the high-temperature compressed air, undergoing combustion at constant pressure. The resulting high-temperature, high-pressure flue gas expands after combustion, enters the turbine region, and drives the power blades to rotate at high speed until it is discharged from the outlet as exhaust gas. This exhaust gas can be reused, for example, by feeding it into a waste heat boiler in a gas turbine generator set to form a combined cycle. For gas turbines, changes in unit load cause changes in turbine power, resulting in continuous changes in the unit's operating state and exhibiting typical large-range non-stationary transient characteristics. To address this characteristic, this invention proposes a hybrid operational status monitoring method combining a knowledge matrix and condition-driven approaches for large-scale non-stationary processes. Figure 1 The diagram shows a flowchart of the operational status monitoring method that combines knowledge matrix and condition-driven approaches according to the present invention. The method of the present invention includes the following steps:

[0049] (1) Obtain N samples of the normal continuous operation process of the gas turbine system. Each sample contains J measured variables, and a two-dimensional data matrix Y(N×J) can be obtained.

[0050] In this example, approximately 3,000 samples were collected from a power plant in Zhejiang Province during normal operation to build the model. The measured variables are as follows: natural gas volume flow rate of the pre-module, natural gas mass flow rate of the pre-module, compressor inlet temperature (representing ambient temperature), compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust bearing generator end temperature, compressor thrust bearing gas turbine end temperature, compressor bearing vibration, compressor side shaft vibration, compressor inlet pressure differential, compressor anti-icing device inlet electric regulating valve position, gas turbine exhaust average temperature, gas turbine side shaft vibration, combustion chamber pressure differential, gas turbine cooling air regulating valve position 1, gas turbine cooling air regulating valve position 2, gas turbine humming, gas turbine speed, gas turbine power, gas turbine second-stage stationary vane ring chamber cooling air pressure, and gas turbine third-stage stationary vane ring chamber cooling air pressure.

[0051] (2) Knowledge-driven selection of conditional indicator variables, specifically including the following sub-steps:

[0052] (2.1) Divide into sub-modules: Based on the mechanistic knowledge and the physical structure of the object, divide the whole object into S sub-modules;

[0053] In this example, the gas turbine is divided into the following three sub-modules: compressor, combustion chamber, and gas turbine.

[0054] (2.2) Establishing the knowledge matrix: For the two-dimensional data matrix Y described in step 1, calculate the knowledge matrix X for any pairwise measured variables X. m X n The correlation coefficients between the data in the m and nth columns of the two-dimensional data matrix Y are:

[0055]

[0056] Where E(·) and D(·) represent the mean and variance of the measured variable in N samples, respectively; corr(X m X n )∈[-1,1] is the measured variable X m X n The correlation coefficient between the two variables is positive, indicating a positive correlation, and negative, indicating a negative correlation. The closer the absolute value of the correlation coefficient is to 1, the stronger the linear correlation between the two variables.

[0057] Based on the calculated correlation coefficients and combined with the mechanistic knowledge of the objects, the correlation between the measured variables can be derived, and thus the knowledge matrix M = [M mn ] J×J And there are:

[0058]

[0059] Among them, Mmn Representative variable X m and X n Whether there is a correlation between the two variables; δ is the correlation coefficient threshold, exceeding which a significant linear correlation can be considered to exist between the two variables; X m ∝X n Represents variable X m and X n There is a correlation in their mechanisms;

[0060] (2.3) Selecting Sub-module Variables: For the S sub-modules divided in step 2.1, based on the knowledge matrix M constructed in step 2.2 and the object mechanism knowledge, select a condition indicator variable X for each sub-module s that can best indicate the process running status of that sub-module. s，c and other related variables X associated with this variable. s，o Together they form a two-dimensional sample matrix of submodule s. J s For the variable dimensions on submodule s;

[0061] (2.4) Selecting the overall conditional indicator variable: Based on the knowledge matrix M constructed in step 2.2 and the object mechanism knowledge, select a conditional indicator variable X that is related to most variables in the knowledge matrix and can indicate the overall process operation status of the object in terms of mechanism. c ;

[0062] (3) Data preprocessing, which includes the following sub-steps:

[0063] (3.1) Calculate the sample difference: For the two-dimensional data matrix X of each sub-module described in step 2.3 s At time t, i.e., the t-th sample Calculate the difference using the following formula:

[0064]

[0065] in, For sample x s，t The first difference, For J s The step involves taking a real number of dimensions x from the sample x. s，t Expand to All time-samples Construct a matrix The sample at the first time step directly uses the original sample value.

[0066] Among them, the condition indicator variable X s，c The matrix is ​​obtained by expanding the samples at all times.

[0067] (3.2) Reconstructing the data matrix: For each sub-module sample matrix X obtained in step 2.3 s , and its condition indicator variable X s，c , Rearrange the samples in ascending order of conditional indicator variable values. Determine the conditional interval σ, divide the conditional indicator variable into K conditional intervals, and construct a data matrix for each conditional interval, forming K static conditional slices. and K dynamic condition slices Where N k，s Let be the number of samples in the k-th condition slice of submodule s, where k = 1, ..., K. Then, mean removal is performed on each condition slice to obtain the standardized condition slice data matrix. and

[0068]

[0069]

[0070] (4) Condition-driven PCA modeling, which includes the following sub-steps:

[0071] (4.1) Establishing the condition slice PCA model: Standardizing the condition slice data matrix for each sub-module The PCA model is established using the following formula:

[0072]

[0073] in, express The corresponding principal components, This represents the PCA transformation matrix.

[0074] (4.2) Determine T 2 The formula for the control limits of the statistical test is as follows:

[0075]

[0076] Among them, Ctrl k，s The submodule s represents the control limits on condition slice k, p is the number of principal elements, and F is the control limit on condition slice k. α (p, N) k，s -p) corresponds to a confidence level of 1-α, a first degree of freedom of p, and a second degree of freedom of N. k，s The F-distribution of -p;

[0077] (4.3) Merging condition slices: Starting from the first condition slice, the next condition slice is combined with the previous condition slice in turn to obtain the condition segment data matrix. Where 'l' represents that the condition segment is composed of 'l' condition pieces. Then, referring to formulas 6 and 7, calculate the PCA model for the condition segment. And control limits Ctrl l，s ;

[0078] (4.4) Determine the breakpoint of the condition segment: Take a constant scaling factor γ and compare the control limits of the condition segment with Ctrl. l，s and the control limits of each condition piece within this condition segment (Ctrl) k，s If Ctrl exists l，s >γ·Ctrl k，s If the condition limit has changed significantly and is no longer close to the control limit of the condition piece, then condition segmentation is required; otherwise, continue splicing condition pieces until condition segmentation is required.

[0079] Assuming that when dividing the condition segment, l = l * Then l * Set the condition segment breakpoint, and set the first l... * Each condition piece is divided into the same condition segment, and these divided condition pieces are removed. The remaining condition pieces are then divided again.

[0080] (4.5) Conditional segment partitioning: Repeat steps 4.3-4.4 until all data is partitioned, finally obtaining C. s One condition segment;

[0081] (4.6) Establish the condition segment PCA model: For the pre-divided C s For each condition segment, refer to formulas 6 and 7 to calculate the PCA model for that condition segment.

[0082] (4.7) Calculate the Bayesian inference distance monitoring index: For each condition segment, establish a Gaussian mixture model (GMM) in the feature space, estimate the GMM parameters using the EM algorithm, and determine the number of Gaussian elements using the FJ algorithm. The probability density function is shown in the following formula:

[0083]

[0084] in, v is a parameter of the probability density function. c Let be the number of Gaussian elements in the feature space of condition segment c. Let be the prior probability of the i-th Gaussian element in condition segment c. Let g be the GMM parameter of the i-th Gaussian element of condition segment c, and g(·) be the probability density function.

[0085] Furthermore, the Bayesian inference distance index for condition segment c is calculated:

[0086]

[0087]

[0088] Where, N c，s Let be the number of samples in condition segment c within submodule s. The posterior probability that sample j belongs to the i-th Gaussian component is calculated based on the probability density function. For T j，c，s Mahalanobis distance to the i-th Gaussian component; T j，c，s express The j-th row; based on the confidence level 1-α, the BID can be calculated using the kernel density estimation method. c，s Control limits Ctrl c，s ;

[0089] (4.8) C is also divided according to the breakpoints of the static condition segment. s By considering each dynamic condition segment and referring to step 4.7, the Bayesian inference distance index of each submodule s on each dynamic condition segment c can be obtained. and control limits

[0090] (5) Determine the weight coefficients of the sub-modules: For the overall condition indicator variable X obtained in step 2.4 c The conditional indicator variable is divided into Z equal intervals, and samples whose overall conditional indicator variable falls within a certain interval are placed into the corresponding sample set, resulting in Z groups of samples. This is combined with the sub-module conditional indicator variable X obtained in step 2.3. s，c Calculate the correlation coefficient for each interval:

[0091]

[0092] in, Let z be the global conditional indicator variable for the samples in interval z. Let z be the conditional indicator variable for the submodule of the sample in interval z.

[0093] Furthermore, the weight coefficient of submodule s when the overall condition indicator variable is in the interval z can be calculated.

[0094]

[0095] Weighting coefficient This represents the degree to which the monitoring results on submodule s contribute to the final result;

[0096] (6) Online monitoring, which includes the following sub-steps:

[0097] (6.1) Obtain the test sample xnew (1×J), and processed according to steps 2.3 and 3.1 to obtain differential samples. and its samples x in each submodule new，s Determine the interval z = z based on the overall conditional indicator variable value of the sample. new And determine the condition segment to which the sample belongs in each model based on the value of the submodule condition indicator variable;

[0098] (6.2) Assume that it belongs to the cth submodule in submodule s. s Given a condition segment, its eigenvalue can be calculated:

[0099]

[0100] (6.3) Calculate the monitoring statistics for the sample in the submodule:

[0101]

[0102]

[0103] (6.4) Based on the interval z new Submodule weight coefficients, calculation of fusion statistics and control limits:

[0104]

[0105]

[0106]

[0107]

[0108] (6.5) Determine the operating status: The gas turbine is considered to be operating under normal conditions if and only if both static and dynamic monitoring indicators are less than their control limits; otherwise, the gas turbine is considered to have a static deviation or dynamic abnormality and further inspection is required.

[0109] First, using the knowledge matrix construction method of this invention, a knowledge matrix and conditional indicator variables are established with a correlation coefficient threshold δ = 0.6, as follows: Figure 2 As shown in (a) and (b), Figure 2 (a) shows the name and corresponding number of the variable being measured. Figure 2 (b) is the knowledge matrix. Figure 3This table presents conditional indicator variables. Taking the compressor component as an example, from a mechanistic perspective, the compressor pressure ratio is a key variable that directly reflects the compressor's operating state, and there is a clear positive correlation between the compressor pressure ratio and the compressor outlet pressure. From a knowledge matrix perspective, the compressor outlet pressure variable is correlated with several other variables; therefore, the compressor outlet pressure (X4) can be selected as the conditional indicator variable for the compressor. Similarly, combining expert knowledge and the constructed knowledge matrix, the gas turbine power (X...) is ultimately selected. 20 ) is used as the overall condition indicator variable, and the combustion chamber pressure difference (X) is selected. 15 ), gas turbine exhaust average temperature (X) 13 These are used as conditional indicator variables for the combustion chamber and the gas turbine, respectively.

[0110] Next, the monitoring method of this invention is used to monitor the operational status of an abnormal process. The value of the constant scaling factor γ is determined by cross-validation. The monitoring results at a value of γ = 1.05 are as follows: Figure 4 As shown in (a) and (b), Figure 4 (a) shows the results of static indicator monitoring. Figure 4 (b) shows the results of dynamic indicator monitoring. In this fault case, the actual time of the fault occurrence was the 3267th sample, from... Figure 4 As shown in (a), the dynamic and static monitoring statistics for the first 3314 normal samples were within the control limits. However, starting from the 3315th sample, the static monitoring statistics began to exceed the limits continuously, with a time delay of 48 samples. The exceeding of the static monitoring statistics indicates a static deviation between the current operating state and the stable operating state in the training data. Upon actual inspection, it was found that the fault was caused by abnormal compressor cooling air pressure due to compressor dirt and other reasons. Figure 4 (c) Comparison results between the method proposed in this invention and traditional monitoring methods. The comparison shows that the detection latency of the method proposed in this invention is significantly lower than that of traditional monitoring methods. Overall, the hybrid monitoring strategy of knowledge matrix and condition-driven approach proposed in this invention can introduce expert knowledge into the field of condition monitoring. Through the condition-driven monitoring strategy, it can distinguish various operating conditions of large-scale non-stationary processes. Finally, based on the condition indicator variable values ​​of each sub-component, the monitoring results of each part are integrated to form an overall monitoring index. This enables more accurate judgment of changes in the overall operating status of the object, effectively improving the timeliness, accuracy, and reliability of actual online monitoring. It helps industrial engineers make accurate judgments on the process operating status of equipment, ensuring the safe and reliable operation of the actual production process.

[0111] The proposed method for monitoring the operational status of gas turbines, which combines knowledge matrix and condition-driven approaches, addresses the structural complexity of large industrial equipment and its wide-ranging non-stationary characteristics. It constructs a knowledge matrix using expert knowledge and correlation analysis, dividing it into sub-modules to reduce model coupling. Then, a condition-driven monitoring method is introduced to divide the data into condition segments and establish multiple monitoring models, thereby solving the monitoring problem of wide-ranging non-stationary processes. Application to actual industrial processes has successfully demonstrated its ability to leverage expert knowledge to guide the establishment and fusion of condition-driven models, improving the timeliness, accuracy, and reliability of online monitoring.

[0112] It should be understood that the present invention is not limited to the gas turbine operation process of the specific embodiments described above. Those skilled in the art can make equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for monitoring the operating state of a gas turbine oriented knowledge condition hybrid drive, characterized by Specifically comprising: The monitoring model is constructed, comprising the following steps: (1) Obtaining the normal continuous operation process of the gas turbine system Each sample contains 10 samples. One measured variable; (2) calculating the correlation coefficient between vectors composed of N samples of any two measured variables , according to the calculated correlation coefficient and combined with the mechanism knowledge of the gas turbine, a knowledge matrix is constructed , wherein the element in the mth row and the nth column of the knowledge matrix represents whether the variables and are associated; 1 indicates association, and 0 indicates no association;​ (3) According to the mechanism knowledge of gas turbine and the physical structure of gas turbine, the whole gas turbine is divided into several sub-modules; and according to the constructed knowledge matrix and the mechanism knowledge of gas turbine, a condition indicating variable which can indicate the process running state to the greatest extent and other related variables associated with the variable are selected in each sub-module to jointly form a two-dimensional sample matrix of the sub-module; meanwhile, according to the constructed knowledge matrix and the mechanism knowledge of gas turbine, a condition indicating variable which can indicate the whole process running state of gas turbine to the greatest extent is selected; ;​​​​​​​​ (4) two-dimensional data matrix of each sub-module , calculate the first-order difference of each sample and the previous sample , combine each sample and its corresponding first-order difference as a new sample , the of all samples at all times constitutes a dynamic two-dimensional sample matrix ; (5) the two-dimensional sample matrix of the sub-module , the dynamic two-dimensional sample matrix , the samples are rearranged in the order of the conditional indicator variable value from small to large and divided into , forming static condition pieces and dynamic condition pieces , wherein is the number of samples in the th static condition piece or dynamic condition piece of the sub-module ; then, each static condition piece and dynamic condition piece is subjected to mean removal processing to obtain the data matrix of the standardized static condition piece and dynamic condition piece and ; the standardized static condition pieces are combined to obtain static condition sections, wherein each static condition section is composed of a plurality of continuous standardized static condition pieces, and the control limit calculated based on the PCA model for each static condition section should be less than the product of the control limit calculated based on the PCA model for each static condition piece in the static condition section and a constant proportion factor ; a monitoring model and control limit for each static condition section are established;​​ According to the breakpoints of the static condition section, a dynamic condition section is also divided, and a monitoring model and a control limit of each dynamic condition section are established. According to the breakpoints of the static condition section, a dynamic condition section is also divided, and a monitoring model and a control limit of each dynamic condition section are established. According to the breakpoints of the static condition section, a dynamic condition section is also divided, (6) dividing the overall condition indicator variable into Z intervals, calculating each sub-module condition indicator variable in each interval z and the overall condition indicator variable correlation coefficient as the weight coefficient of each sub-module when the overall condition indicator variable is in interval z ; The online monitoring comprises the following steps: Real-time acquisition of the sample to be measured and processing to obtain the first-order difference and its sample on each submodule ; according to the overall condition indicator value of the sample, determine the interval to which it belongs , and according to the submodule condition indicator value of the sample, determine the static condition segment and the dynamic condition segment to which it belongs, and calculate the monitoring index and the control limit of the current sample to be measured according to the monitoring model, the control limit of the static condition segment and the dynamic condition segment to which it belongs, and the weight coefficient of the interval ; only when both the static and dynamic monitoring indexes are less than their control limits, it is considered that the gas turbine is running in a normal state; otherwise, it is considered that the gas turbine has an abnormality.

2. The method of claim 1, wherein, The measured variables comprise multiple of the following: pre-module natural gas volume flow, pre-module natural gas mass flow, compressor inlet temperature, compressor outlet pressure, compressor outlet temperature, compressor bearing temperature, compressor thrust pad bearing generator end temperature, compressor thrust pad bearing turbine end temperature, compressor bearing vibration, compressor side large shaft vibration, compressor inlet duct differential pressure, compressor anti-icing device inlet electric regulating valve position, turbine exhaust average temperature, turbine side large shaft vibration, combustion chamber differential pressure, turbine cooling air regulating valve position 1, turbine cooling air regulating valve position 2, turbine humming, turbine speed, turbine power, turbine 2nd stage static blade holding ring cavity cooling air pressure, turbine 3rd stage static blade holding ring cavity cooling air pressure.

3. The method of claim 1, wherein, The whole gas turbine is divided into The whole gas turbine is divided into The overall gas turbine is divided into a compressor, a combustion chamber and a turbine.

4. The method of claim 1, wherein, The monitoring model and control limit of each static condition section are established The specific process is as follows: establishing a PCA model for each static condition section , ; wherein is a data matrix of the s-th module and the c-th static condition section, denotes the corresponding principal component, denotes the corresponding PCA transformation matrix of the s-th module and the c-th static condition section; For each static condition segment, a Gaussian Mixture Model is built in the feature space and the number of Gaussians is determined using the F-J algorithm The GMM parameters are then estimated using the EM algorithm , The GMM parameters for the static condition segment 's th Gaussian component are estimated , The monitoring model for the static condition segment is obtained PCA model based on static condition segments and GMM parameter construction of probability density function : ； wherein is a parameter of the probability density function, is the prior probability of the th Gaussian component of the conditional segment is the prior probability of the th Gaussian component of the conditional segment is the GMM parameter of the th Gaussian component of the conditional segment Computing the static condition segment Bayesian inference distance metrics: ； ； in, For submodules middle condition section The number of samples, For the sample Belongs to the The posterior probability of each Gaussian component for To the Mahalanobis distance with Gaussian components; express The j-th row; based on confidence level The kernel density estimation method was used to calculate control limits .

5. The method of claim 4, wherein, The sub-module condition indicator value according to the sample determines the static condition section and the dynamic condition section to which it belongs, and calculates the monitoring index and the control limit of the current sample to be tested according to the weight coefficient of the monitoring model, the control limit and the interval of the static condition section and the dynamic condition section to which it belongs , as follows: The monitoring statistics of the to-be-measured sample on the sub-modules are calculated: ； ； Wherein, is the conditional segment serial number of the sample to be tested in the submodule s, , represents the monitoring model corresponding to the conditional segment, , represents the characteristic value of the sample to be tested, specifically: ; According to the sub-module weight coefficient of the interval , the static monitoring index , the dynamic monitoring index and the corresponding control limit , are calculated. ； ； ； 。

Images