A data-driven anomaly detection method based on flight data recorder (FDR) data

Abstract

The present invention provides a data-driven anomaly detection method based on flight data. The method includes: establishing an engine high-temperature component data model based on the extreme learning machine algorithm; determining a residual sequence based on the engine high-temperature component data model; obtaining a residual threshold band based on the residual sequence; and performing anomaly detection on the engine high-temperature components based on the residual threshold band. The present invention comprehensively considers the influences of multiple variables and multiple factors, takes the working conditions as input variables, conducts in-depth mathematical analysis on the engine operation mechanism, selects multiple variables for data modeling, divides the data set into a training set, a validation set, and a test set to improve the generalization ability of the model. Moreover, the ELM with fast learning speed and strong generalization ability is selected as the modeling means, and finally, a high-precision and strong-sensitivity gas turbine high-temperature component detection algorithm is designed.

Description

Technical Field

[0001] The invention belongs to the technical field of helicopter fault detection, and in particular relates to a data-driven anomaly detection method based on flight parameter data. Background Art

[0002] The current detection methods for high-temperature engine components (including combustion chamber flame tubes, fuel nozzles, first-stage turbine guide vanes, rotor blades, etc.) cannot meet actual requirements.

[0003] The main problems in current related research are as follows: Since the information on the trend of turbine exhaust temperature changes has not been effectively utilized, the anomaly detection and diagnosis algorithm based on the maximum temperature or temperature difference has serious hysteresis and cannot provide real-time warning to prevent problems before they occur.

[0004] However, since this type of method is easy to operate and implement, and can detect over-temperature or uneven exhaust temperature distribution to a certain extent, it currently exists in most existing performance detection systems. In addition, the operating conditions and environment still have a great impact on the detection algorithm, and the "false alarm rate" and "missing alarm rate" of the method are more sensitive to different operating conditions of the gas turbine, and the generalization ability still needs to be improved.

[0005] In summary, in order to have stronger practical significance, the present invention comprehensively considers the influence of multiple variables and factors, and designs a high-precision and highly sensitive gas turbine high-temperature component detection algorithm, so as to avoid the "after-the-fact diagnosis" phenomenon as much as possible and realize an effective early warning function. Summary of the invention

[0006] In view of the above technical problems, the present invention provides a data-driven anomaly detection method based on flight parameter data, the method comprising:

[0007] Establish a data model for high-temperature engine components based on the extreme learning machine algorithm;

[0008] Determining a residual sequence based on the engine high temperature component data model;

[0009] A residual threshold band is obtained based on the residual sequence.

[0010] Preferably, the establishing of the engine high temperature component data model based on the extreme learning machine algorithm comprises:

[0011] The average exhaust temperature at the turbine outlet, the air inlet temperature and the gas generator outlet pressure are selected as the input variables for modeling, and the exhaust temperature at the turbine measuring point outlet is selected as the output variable;

[0012] Dividing the input variables and the output variables into a training set, a validation set and a test set according to a preset ratio;

[0013] Determine the structure of the extreme learning machine based on the training set and the Schwarz criterion;

[0014] Verify the structure of the extreme learning machine based on the validation set to determine the second engine high-temperature component data model.

[0015] Preferably, after verifying the structure of the extreme learning machine based on the validation set to determine the second engine high-temperature component data model, it further includes:

[0016] Test the second engine high-temperature component data model based on the test set.

[0017] Preferably, determining the structure of the extreme learning machine based on the training set and the Schwarz criterion includes:

[0018] Train the first engine high-temperature data model based on the number of hidden layer nodes of the extreme learning machine structure to obtain the target Schwarz criterion number;

[0019] Determine the number of hidden layer nodes corresponding to the target Schwarz criterion number.

[0020] Preferably, determining the residual sequence based on the engine high-temperature component data model includes:

[0021] Obtain the predicted value of the output variable based on the value of the input variable and the engine high-temperature component data model;

[0022] Obtain the residual sequence based on the measured value of the output variable and the predicted value of the output variable; wherein, the measured value of the output variable is the value recorded by the flight parameter recording system.

[0023] Preferably, obtaining the residual threshold band based on the residual sequence includes:

[0024] Conduct a normality test on the residual sequence;

[0025] If the values of the residual sequence conform to a normal distribution, determine the residual threshold band based on the box plot and the residual sequence.

[0026] Preferably, conducting the normality test on the residual sequence includes:

[0027] Conduct a normality test on the residual sequence based on the Q-Q plot.

[0028] Preferably, the method further includes:

[0029] Conduct anomaly detection on the engine high-temperature components based on the residual threshold band.

[0030] The beneficial technical effects of the present invention:

[0031] The present invention comprehensively considers the influence of multiple variables and multiple factors, takes the operating conditions as input variables, conducts in-depth mathematical analysis on the engine operation mechanism, selects multiple variables for data modeling, divides the data set into a training set, a validation set, and a test set to improve the generalization ability of the model, and selects ELM with fast learning speed and strong generalization ability as a modeling means, and finally designs a high-precision and highly sensitive gas turbine high-temperature component detection algorithm, thereby avoiding the "after-the-fact diagnosis" phenomenon as much as possible and realizing an effective early warning function. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a schematic diagram of the structure of an engine provided by an embodiment of the present invention;

[0033] Figure 2 It is a box plot provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] See also Figure 1-2 The present invention proposes a data-driven abnormality detection method for high-temperature components of a gas turbine with a fusion mechanism, aiming to improve detection accuracy and detect faults early.

[0035] The present invention proposes an abnormality detection method for high-temperature components of an engine based on an extreme learning machine, which mainly includes two modules: establishing a data model of high-temperature components of a gas turbine based on ELM, and designing an abnormality detection method for high-temperature components of a gas turbine based on ELM residual analysis.

[0036] Data model of high temperature components of gas turbine based on ELM

[0037] Establishing a data model for high-temperature components of a gas turbine includes four parts: algorithm selection, feature selection, data partitioning, and model building.

[0038] Among them, the calculation process of the extreme learning machine algorithm (ELM) is:

[0039] The extreme learning machine algorithm can effectively avoid the repeated iteration process in the traditional neural network training process. Only by determining the number of neurons in the hidden layer and the activation function of the hidden layer and neurons, ELM can calculate the parameters between the hidden layer and the output layer through the randomly generated weights w and bias b between the input layer and the hidden layer. Therefore, the present invention selects ELM to establish the data model. Specifically, the learning algorithm of ELM can be summarized into the following steps:

[0040] (1) Determine the number of neurons in the hidden layer, randomly set the connection weight w between the input layer and the hidden layer and the bias b of the neurons in the hidden layer;

[0041] (2) Select an infinitely differentiable function as the activation function of the hidden layer neurons, and then calculate the hidden layer output matrix H;

[0042] (3) Calculate the weights of the output layer.

[0043] Input Feature Selection of the ELM Model Based on Gas Turbine Mechanism

[0044] According to the variable representation, assume that in the multi-dimensional space X, the exhaust gas temperature T4 measured by the thermocouple is a point.

[0045] T4 = [T 4,1 , T 4,2 ,..., T 4,i ..., T 4,n (1)

[0046] Due to the influence of interference, the distance of the data sample T4 in the X space is too large. Therefore, the data sample must be appropriately transformed from the X space to the Y space. In the multi-dimensional space Y, the internal distance is small, and normal operation and abnormal operation are easy to distinguish.

[0047] Since the energy absorbed by the gas in different combustion chambers is different, the readings of the exhaust gas thermocouples are also different. The inherent structural differences between different combustion chambers are considered to be caused by processing and installation errors. And this inherent structural difference will also cause differences in the process of the gas absorbing energy.

[0048] Assume

[0049] Q i = μ i Q (2)

[0050] Where Q is the energy absorbed by the gas in the combustion chamber, and Q i describes the energy absorbed by the gas in the i-th combustion cylinder.

[0051] Therefore, the coefficient μ i is related to the energy distribution of each combustion cylinder and characterizes the inherent structural information of each combustion cylinder. In addition,

[0052] Assume that the gas flow is continuous. According to the thermodynamics theory, the following equation is obtained:

[0053]

[0054]

[0055] C p,3 T 3,i = C p,2 T2 + μ i Q (5)

[0056] Based on equations (3)-(5), the following equation is obtained:

[0057]

[0058] Assume that the temperature measured by the i-th thermocouple corresponds to the i-th combustion chamber. Therefore, Equation (7) can be obtained and rewritten as:

[0059]

[0060]

[0061] According to Equations (7) and (8), the thermocouple temperature can be expressed by the following equation:

[0062]

[0063] where C p,2 , C p,3 and k are constants. Equation (10) can be derived:

[0064] T 4,i = μ i nT 4,avg + f(T1, P2)(1 - μ i n) (10)

[0065] Therefore, in order to better match the operating mechanism of the gas turbine and reduce the degree of modeling simplification. From Equation (10), the present invention selects the average exhaust temperature at the turbine outlet, the air inlet temperature, and the gas generator outlet pressure as the input variables for subsequent modeling.

[0066] ELM Model Establishment

[0067] The relevant characteristic parameter data sources recorded in the flight parameters are divided into a training set, a validation set, and a test set in a ratio of 70%, 15%, and 15%, so as to ensure the generalization ability of the established model. Through the mechanism analysis of the engine, the average exhaust temperature at the turbine outlet, the air inlet temperature, and the gas generator outlet pressure are determined as the model input variables. Based on the Schwarz criterion, the model accuracy and complexity are comprehensively evaluated, and the number of hidden layer nodes of the ELM network is determined to realize the selection of the model structure.

[0068] The Schwarz criterion, abbreviated as the SC criterion, is a method for determining the maximum lag period. The expression for its test is:

[0069]

[0070] where L is the maximum likelihood value of the parameter to be estimated, k is the number of explanatory variables in the model, and T is the sample size. The principle of this method is to minimize the SC value during the process of increasing the k value. When its value is the smallest, it is the length of the optimal lag distribution.

[0071] The present invention selects the ELM model using the SC criterion. With the goal of minimizing the SC criterion number, the model is trained by continuously increasing the number of hidden layer nodes, and the number of hidden layer nodes corresponding to the minimum SC criterion number is selected as the number of hidden layer nodes of the ELM model for each exhaust temperature measurement point.

[0072] Abnormal Detection Method for Gas Turbine High-Temperature Components Based on ELM Residual Analysis

[0073] Regarding the problem of abnormal detection of gas turbine high-temperature components studied in the present invention, by analyzing the residual sequence generated during the ELM modeling process, a residual threshold band is obtained and used as a criterion for abnormal detection. The main contents include: performing a normality test and correlation analysis on the residuals, and obtaining the abnormal detection threshold band.

[0074] Normality Test of Residuals

[0075] To verify the normality of the training set residuals, a Q-Q plot is used to test them. In statistics, a probability distribution can be represented by a Q-Q plot (Q represents quantile), which provides a method for representing a probability distribution, that is, the graphs of the two quantiles to be compared are placed together for analysis and comparison. If the obtained Q-Q plot is close to this straight line, it indicates that the two distributions to be verified have the same distribution.

[0076] Design of Abnormal Detection Strategy Based on ELM Residual Box Plot

[0077] A box plot is a statistical graph used to display the dispersion of a set of data. Its drawing uses common statistical quantities and can provide key information about the location and dispersion of the data. It is named because of its shape like a box. Figure 2 Indicates the meaning of each line in the figure, and the concept of quantile value (number) is applied. It mainly includes six data nodes. Arrange a set of data from largest to smallest, and calculate its upper edge, upper quartile Q2, median, lower quartile Q1, lower edge, and outliers respectively.

[0078] The box plot provides a criterion for identifying outliers: Outliers are defined as values less than Q1 - 1.5IQR or greater than Q1 - 1.5IQR, where IQR = Q2 - Q1. This criterion comes from empirical judgment and performs well in dealing with data that requires special attention. Compared with the traditional 3σ rule or z-score method based on normal distribution, which assumes that the data follows a normal distribution, but the actual data often does not strictly follow a normal distribution.

[0079] The present invention establishes a data model of high-temperature components of a gas turbine based on ELM, comprehensively analyzes and considers the operation mechanism of the gas turbine and the complexity of modeling. The present invention first performs feature selection, and selects the average exhaust temperature at the turbine outlet, the air inlet temperature, and the gas generator outlet pressure as input variables for subsequent modeling. Then, an ELM with fast learning speed and strong generalization ability is selected for modeling, and the data source is divided into a training set, a validation set, and a test set (70%, 15%, and 15%), and the average exhaust temperature at the turbine outlet, the air inlet temperature, and the gas generator outlet pressure are used as model input variables. The model accuracy and complexity are comprehensively considered to determine the selection of the model structure, and the high-temperature component data model based on ELM is completed.

[0080] The abnormality detection method for high temperature components of a combustion engine by ELM residual analysis of the present invention performs normality test and correlation analysis on the residual sequence generated in the ELM modeling process to obtain an evaluation index (residual threshold band) for abnormality detection.

[0081] Compared with the current detection methods for high-temperature engine parts, the diagnostic algorithms of the present invention have serious hysteresis, high "false alarm rate" and "missing alarm rate", and low generalization ability. The present invention comprehensively considers the influence of multiple variables and multiple factors, takes the working conditions as input variables; conducts in-depth mathematical analysis of the engine operation mechanism, selects multiple variables for data modeling; divides the data set into training set, verification set, and test set to improve the generalization ability of the model; and selects ELM with fast learning speed and strong generalization ability as a modeling means, and finally designs a high-precision and highly sensitive gas turbine high-temperature component detection algorithm, thereby avoiding the "after-the-fact diagnosis" phenomenon as much as possible and realizing an effective early warning function.

Claims

1. A data-driven anomaly detection method based on flight data recorder (FDR) data, characterized in that The method includes: Establishing a data model of engine high-temperature components based on the extreme learning machine algorithm; Determining a residual sequence based on the data model of engine high-temperature components; Obtaining a residual threshold band based on the residual sequence; Among them, establishing a data model of engine high-temperature components based on the extreme learning machine algorithm includes: According to the variable representation, assume that in a multi-dimensional space the exhaust gas temperature T4 measured by the thermocouple is a point; Due to the influence of interference, the distance of data sample T4 in X space is too large; therefore, the data sample is converted from X space to space. In the multi-dimensional space Y, the internal distance is small, and normal operations and abnormal operations are easy to distinguish; Since the energy absorbed by the gas in different combustion chambers is different, the readings of the exhaust thermocouples are also different; the inherent structural differences between different combustion chambers are considered to be caused by processing and installation errors; and this inherent structural difference will also cause differences in the process of gas absorbing energy; Assume where Q is the energy absorbed by the gas in the combustion chamber, and Q i describes the energy absorbed by the gas in the i-th combustion cylinder; Therefore, the coefficient µ i is related to the energy distribution of each combustion chamber and characterizes the inherent structural information of each combustion chamber; in addition, ; Assume that the gas flow is continuous. According to the thermodynamics theory, the following equation is obtained: Based on Equation 3-5, the following equation is obtained: Assume that the temperature measured by the i-th thermocouple corresponds to the i-th combustion chamber; therefore, Equation 7 can be obtained and rewritten as Equation 8: According to Equations 7-8, the thermocouple temperature can be expressed by the following equation: where C p,2 , C p,3 and k are constants, and the equation 10 can be derived as follows: Select the average exhaust temperature at the turbine outlet, the air inlet temperature, and the gas generator outlet pressure as the input variables for modeling, and select the exhaust temperature at the turbine measurement point outlet as the output variable; Dividing the input variables and the output variables into a training set, a validation set, and a test set according to a preset ratio; Determining the extreme learning machine structure based on the training set and the Schwarz criterion; Validating the extreme learning machine structure based on the validation set to determine the second data model of engine high-temperature components.

2. The method according to claim 1, wherein After validating the extreme learning machine structure based on the validation set to determine the second data model of engine high-temperature components, it further includes: Testing the second data model of engine high-temperature components based on the test set.

3. The method according to claim 1, wherein Determining the extreme learning machine structure based on the training set and the Schwarz criterion includes: Training the first engine high-temperature data model based on the number of hidden layer nodes of the extreme learning machine structure to obtain the target Schwarz criterion number; Determining the number of hidden layer nodes corresponding to the target Schwarz criterion number.

4. The method according to claim 1, wherein Determining the residual sequence based on the data model of engine high-temperature components includes: Obtaining the predicted value of the output variable based on the values of the input variables and the data model of engine high-temperature components; Obtaining the residual sequence based on the measured value of the output variable and the predicted value of the output variable; among them, the measured value of the output variable is the value recorded by the flight parameter recording system.

5. The method according to claim 1, characterized in that Obtaining the residual threshold band based on the residual sequence includes: Conducting a normality test on the residual sequence; If the values of the residual sequence conform to the normal distribution, determining the residual threshold band based on the box plot and the residual sequence.

6. The method according to claim 1, wherein Conducting a normality test on the residual sequence includes: Conducting a normality test on the residual sequence based on the Q-Q plot.

7. The method according to claim 1, wherein The method further includes: Conducting anomaly detection on engine high-temperature components based on the residual threshold band.

Images