An aero-engine state monitoring method based on information increment

Abstract

The application discloses an aero-engine state monitoring method based on information increment. According to the information increment of the window data matrix, all the information gain indexes of the mutations are recognized by using the window movement, the switching points of the task stages are further determined, and the task stage division of the single running process of the aero-engine is completed. The adjustable parameters are adjusted according to the correlation requirement, the accurate division of the task stages is realized, the task stages with high similarity in different running data are rearranged into one task stage by using the similarity analysis, and the monitoring sub-models are respectively established for each task stage, so that the modeling problems of the task stages with different lengths and the different change trends of the task stages in each running process of the aero-engine are solved, the fault monitoring is more effective, and the method has certain practical significance.

Description

Technical Field

[0001] This invention relates to the field of aircraft engine condition monitoring technology, and in particular to an aircraft engine condition monitoring method based on information increment. Background Technology

[0002] Throughout the flight of a civil aircraft, the aero-engine system needs to utilize different variable information, control laws, management rules, and even key subsystems to complete the tasks of each flight phase (takeoff, climb, cruise, descent, approach, and landing). This means that the operational characteristics and variable features of different flight phases differ significantly. Compared to a single-mission system (stationary process), the operation of the aero-engine system is more complex due to the differences in objectives between different missions. This complexity is mainly manifested in its multi-phase characteristics, the unequal duration of mission phases, and the differences in the trends of change during operation.

[0003] The operation of aero-engines exhibits multi-stage characteristics, meaning that process variables change regularly following the process operation or the changes in process mechanism characteristics. If a single monitoring model is used, the differences between mission stages will be weakened, thus affecting the accuracy of the model. Therefore, in order to ensure the stability and safety of aero-engine operation, it is necessary to correctly, reasonably, and reliably identify and divide each mission stage to ensure that subsequent monitoring and fault diagnosis meet the requirements of high performance. Currently, most methods have been studied for multi-stage characteristics and unequal stage durations, such as stage identification and online monitoring of non-uniform intermittent processes [1], using the k-nearest neighbor rule of variable moving window to construct pseudo-time slices for unequal stage division and process monitoring [2], and stage division of multi-stage intermittent processes based on information increment matrix [3]. All the above methods have good repeatability, so the change trends in different operation processes are very similar regardless of whether they have the same duration. However, for aero-engine systems, the trend of mission stage changes may vary greatly in each operation. For example, some flight mission stages (climb and approach) may consist of multiple discontinuous flight mission sub-segments. Therefore, this essential characteristic makes it difficult to establish aero-engine monitoring models, which in turn affects the performance of monitoring.

[0004] [1]Guo, R., & Jin, Y. (2019). Phase Identification and Online Monitoring for the Uneven Batch Processes. IEEE Access, 7, 81351-81363.

[0005] [2] Zhang, S., Zhao, C., Wang, S., & Wang, F. (2017). Pseudo Time-SliceConstruction Using a Variable Moving Window k Nearest Neighbor Rule for Sequential Uneven Phase Division and Batch Process Monitoring. Industrial & Engineering Chemistry Research, 56 (3), 728-740.

[0006] [3] Li Zheng, Wang Pu, Gao Xuejin, et al. Quality prediction of multi-stage intermittent processes based on information increment matrix [J]. Journal of Chemical Industry and Engineering (China), 2018, 69(12):5164-5172. Summary of the Invention

[0007] In view of the current state of technology and the multi-stage characteristics of aero-engine operation, the unequal length of mission stages, and the differences in changing trends, this invention proposes an aero-engine condition monitoring method based on information increment.

[0008] The technical solution adopted in this invention is as follows: A method for monitoring the condition of an aero-engine based on information increment includes the following steps:

[0009] S1: The window data matrix acquisition module uses a window to select single-run data from the historical operation data of the aero-engine to obtain the window data matrix.

[0010] S2: The task phase division module calculates the information gain index of the window data matrix using the information increment method, and then determines the switching point of the task phase by identifying the information gain index of abrupt changes, thus obtaining the divided task phases.

[0011] The task phase division includes the following steps:

[0012] S21: The information gain index of the window data matrix is ​​calculated using the information increment method. First, the window data matrix... Preprocessing is performed to obtain The calculation formula is as follows:

[0013]

[0014]

[0015] In equation (1), b is The mean vector, l = [1, 1, ..., 1] T∈R w×1 k = 1, 2, ..., K i -w+1;

[0016] Then the preprocessed window data matrix Calculate the covariance matrix R k The calculation formula is as follows:

[0017]

[0018] Then use the two adjacent covariance matrices R h and R h+1 Calculate the information increment matrix D h The calculation formula is as follows:

[0019] D h =R h+1 -R h ,(1≤h≤K i -w) (3)

[0020] Finally, based on the information increment matrix D h The information gain index δ is calculated. h The calculation formula is as follows:

[0021]

[0022] In equation (4), D h (a,b) represents the information increment matrix D. h The element in row a, column b, 1 ≤ h ≤ K i -w, a = 1, 2, ..., J, b = 1, 2, ..., J, where J is the number of process variables.

[0023] S22: Use a window to apply the information increment index δ h The selection is performed, and the window selects L consecutive information increment indices δ. h The window updates its information increment index by moving backward by one information increment index each time, and then sequentially judges each information gain index δ within the window. h Whether the information gain index of the mutation is satisfied is determined by the following formula:

[0024] δ h ≥mean(d)+μ·std(d),(d≤h≤L+d-1) (5)

[0025] In equation (5), mean(d) and std(d) are the information gain indices δ for all information gain indices within the d-th window, respectively. h The mean and standard deviation, μ is an adjustable parameter, d = 1, 2, ..., K i -wL; when the information gain index δ hWhen formula (5) is satisfied, the information gain index δ h It is the information gain index of the mutation, and the information gain index of all mutations is identified according to formula (5).

[0026] S23: The window data matrix corresponding to the information increment index of all the mutations The final sample point is used as the switching point to obtain all switching points. The running data is then divided into V based on all switching points. i Each task phase, V i The data for each task phase is denoted as

[0027] S3: The task phase optimization module adjusts the adjustable parameter values ​​by judging whether the correlation coefficient between the centroids of adjacent task phases in the divided task phases meets the correlation requirements, so as to obtain the accurately divided task phases.

[0028] The precise division of the task phases involves the following steps:

[0029] S31: Calculate the partition of the running data to obtain V i The correlation coefficient between the centroids of data from two adjacent task stages is calculated using the following formula:

[0030]

[0031] In equation (6), and It is the centroid of two adjacent task phases in the data of the i-th run. yes and The correlation coefficient, yes and covariance, and They are and The variance, v = 1, 2, ..., V i .

[0032] S32: Sequentially determine the correlation coefficient between the centroids of the data from two adjacent task stages. Does it meet the relevance requirements? If satisfied, the adjustable parameter μ is updated according to μ = μ + 0.1, and formula (5) is returned; otherwise, the precisely divided task stages are obtained, and the running data is precisely divided to obtain C. i Each task phase, C i The data for each task phase is denoted as

[0033] S4: Through the window data matrix acquisition module, task phase division module and task phase optimization module, the historical operation data of the aero-engine is accurately divided into each operation data to obtain the task phase data in all operation data.

[0034] S5: The task phase rearrangement module calculates the similarity of task phase data in all running data, rearranges the task phases with high similarity in different running data into one task phase, and determines each task phase after rearrangement.

[0035] The task phase reordering includes the following steps:

[0036] S51: Arrange the data from I runs from top to bottom with the same variable dimensions to obtain a matrix. Calculate matrix The mean and standard deviation of the data, where the mean of all data for the j-th variable is... The calculation formula is The standard deviation s of all data for the j-th variable j The calculation formula is i = 1, 2, ..., I, k t =1,2,…,K t J is the number of process variables, K r K is the total number of sample points in the I-run data. i It is the number of sample points in the i-th run.

[0037] S52: According to the matrix mean and standard deviation s j Standardize the data for each task phase in all runtime data, where the data for the c-th task phase of the i-th runtime data is... The data for the standardized task phase is obtained through standardization. k c The j-th variable data of each sample point The standardized calculation formula is as follows:

[0038]

[0039] In equation (7), c = 1, 2, ..., C i , It is the number of sample points in the c-th task stage of the i-th run, j = 1, 2, ..., J.

[0040] S53: Calculate any two standardized task phase data and The Euclidean distance between them is used to obtain the Euclidean distance matrix D. ic,nt (P×Q), i=1,2,…,I, c=1,2,…,C i , n=1,2,…,I, t=1,2,…,C n t≠c, n≠i, P and Q are respectively and The number of sample points, of which data any sample point and data any sample point Euclidean distance between them p,q The calculation formula is as follows:

[0041]

[0042] In equation (8), p = 1, 2, ..., P, q = 1, 2, ..., Q;

[0043] According to the Euclidean distance matrix D ic,nt (P×Q) defines the similarity S. ic,nt The formula is as follows:

[0044]

[0045] In equation (9), D ic,nt (p,q) represents the Euclidean distance matrix D. ic,nt The element in row p and column q.

[0046] S54: For all running data, the standardized task stage data are similar to each other according to formulas (8) and (9). The task stage data with high similarity are expanded as variables and rearranged into a task stage to finally obtain C. * Each task phase;

[0047] S6: The model building modules are C * For each task phase, a PCA monitoring sub-model is constructed to calculate the monitoring statistics and control limits for each task phase.

[0048] S7: The online monitoring module determines whether the currently running online data is in a fault state based on the constructed PCA monitoring sub-model.

[0049] The steps to determine whether the currently running online data is in a fault state are as follows:

[0050] S71: Obtain the current running data online, and use the mean and standard deviation in formula (7) to standardize the current running data to obtain the standardized current data.

[0051] S72: Use formulas (8) and (9) to determine the task stage to which the standardized current running data belongs.

[0052] S73: Calculate the standardized monitoring statistics of the current operating data based on the PCA monitoring sub-model of the task phase to which it belongs.

[0053] S74: Compare the monitoring statistics of the current running data with the control limits of the task stage to which it belongs. If the monitoring statistics of the current running data do not exceed the control limits of the task stage to which it belongs, it means that the current running data is normal; otherwise, it is determined that the current running data is in a fault state.

[0054] Compared with existing technologies, the advantages of this invention are as follows: Based on the information increment of the window data matrix, window movement is used to identify the information gain index of all abrupt changes, further determining the switching points of the mission phases and completing the division of mission phases for a single operation of the aero-engine; adjustable parameters are adjusted according to correlation requirements to achieve precise division of mission phases; similarity analysis is used to rearrange mission phases with high similarity in different operational data into a single mission phase, and a monitoring sub-model is established for each mission phase, solving the modeling problems of unequal mission phase lengths and differences in the changing trends of mission phases during each aero-engine operation, thus enabling more effective fault monitoring and possessing certain practical significance. Attached Figure Description

[0055] Figure 1 This is a flowchart of an aero-engine condition monitoring method based on information increment according to the present invention;

[0056] Figure 2 This is a schematic diagram illustrating the principle of window movement in this invention;

[0057] Figure 3 The graph shows the trend of the information increment index δh when five consecutive sample points are selected to form a window data matrix in this invention.

[0058] Figure 4 The graph shows the trend of the information increment index δh when 15 consecutive sample points are selected to form a window data matrix for the present invention.

[0059] Figure 5 The graph shows the trend of the information increment index δh when 30 consecutive sample points are selected to form a window data matrix for the present invention.

[0060] Figure 6 This is a diagram showing the task stage division results under different adjustable parameter μ values ​​of the present invention;

[0061] Figure 7 This is a schematic diagram illustrating the operational data of the present invention, which has unequal task phase lengths and different trends in task phase changes. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0063] In this embodiment, the historical operation data of the aero-engine consists of the operation data of four different flights of a certain type of aero-engine from Shanghai to Beijing, Paris to Shanghai, and Anchorage to Beijing, as shown in Table 1. The flight duration and the number of sample points are different for each flight. In this embodiment, 14 process variables are selected for monitoring, as shown in Table 2.

[0064] Table 1 Explanation of Historical Operating Data for Aero Engines

[0065] Serial Number Number of sample points 1 7024 2 8983 3 6741 4 8302

[0066] Table 2. Description of Process Variables

[0067]

[0068] The specific steps for condition monitoring of this type of aero-engine are as follows:

[0069] S1: The window data matrix acquisition module uses a window to select data from each run in the historical operation data of the aero-engine to obtain the window data matrix.

[0070] In this embodiment, each operation data in the historical operation data of the aero-engine is recorded as a two-dimensional matrix. in, i = 1, 2, ..., 4, K i J is the number of sample points in the i-th run, and J is the number of process variables, J = 1, 2, ..., 14. Data from each run is processed using a window. Select 15 consecutive sample points within a window, and then construct a window data matrix from these 15 consecutive sample points, as shown below. Figure 2 As shown, the window updates its data matrix by moving one sample point backward each time, ultimately obtaining K. i -14 window data matrices, K i -14 window data matrices are denoted as in, This represents the data matrix of the k-th window. k = 1, 2, ..., K i -14, and These are the initial and final sample points of the window data matrix, respectively.

[0071] S2: The task phase division module calculates the information gain index of the window data matrix using the information increment method, and then determines the switching point of the task phase by identifying the information gain index of abrupt changes, thus obtaining the divided task phases.

[0072] In this embodiment, the task phase division includes the following steps:

[0073] S21: The information gain index is calculated using the information increment method for the window data matrix. First, the window data matrix... Preprocessing is performed to obtain The calculation formula is as follows:

[0074]

[0075]

[0076] In equation (1), b is The mean vector, l = [1, 1, ..., 1] T ∈R w×1 k = 1, 2, ..., k i -14;

[0077] Then the preprocessed window data matrix Calculate the covariance matrix R k The calculation formula is as follows:

[0078]

[0079] Then use the two adjacent covariance matrices R h and R h+1 Calculate the information increment matrix D h The calculation formula is as follows:

[0080] D h =R h+1 -R h ,(1≤h≤K i -15) (3)

[0081] Finally, based on the information increment matrix D h The information gain index δ is calculated. h The calculation formula is as follows:

[0082]

[0083] In equation (4), D h (a,b) represents the information increment matrix D. h The element in row a, column b, 1 ≤ h ≤ K i -15, a=1,2…,14, b=1,2…,14,

[0084] Figure 3 , 4 As shown in Figure 5, the first 1000 information increments δ of the data from the first run in this embodiment are used as an example. h For example, the information increment index δ was analyzed when the window data matrix was constructed by selecting 5, 15, and 30 sample points respectively. h The changing trend reveals that as the number of sample points selected in the window increases, a sudden change in the information increment index δ gradually emerges. h After extensive experimentation, considering both the impact of an excessively large number of sample points in the window on sensitivity to changes and the impact of insufficient window information on reliability, this embodiment sets the window to select 15 sample points.

[0085] S22: As the task phase changes, the information increment index will change significantly. Within the same task phase, the information increment index δ h The information increment index δ changes smoothly within a certain range, and when entering the next task phase, it increases. h Significant changes will occur as we move into the next task phase, with the information increment index δ... h It will then change smoothly again within another range, indicating that if the information increment index δ h A large value does not necessarily indicate a sudden change in the information gain index; it may simply mean strong dynamic characteristics within the current task phase. Therefore, this invention uses a window to measure the information increment index δ. h Select 23 consecutive information increment indices δ in the window. h The window updates its information increment index by moving backward by one information increment index each time, and then sequentially judges each information gain index δ within the window. h Whether the information gain index of the mutation is satisfied is determined by the following formula:

[0086] δ h ≥mean(d)+μ·std(d),(d≤h≤22+d) (5)

[0087] In equation (5), mean(d) and std(d) are the information gain indices δ for all information gain indices within the d-th window, respectively. h The mean and standard deviation, μ is an adjustable parameter, d = 1, 2, ..., K i -38; when the information gain index δ h When formula (5) is satisfied, the information gain index δ h It is the information gain index of the mutation, and the information gain index of all mutations is identified according to formula (5).

[0088] S23: Window data matrix corresponding to the information increment index of all mutations The final sample point is used as the switching point to obtain all switching points. The running data is then divided into V based on all switching points. i Each task phase, V i The data for each task phase is denoted as

[0089] S3: The task phase optimization module adjusts the adjustable parameter values ​​by judging whether the correlation coefficient between the centroids of adjacent task phases in the divided task phases meets the correlation requirements, so as to obtain the accurately divided task phases.

[0090] In this embodiment, the precise division of task phases involves the following steps:

[0091] S31: Calculate the partitioning of runtime data to obtain V i The correlation coefficient between the centroids of data from two adjacent task stages is calculated using the following formula:

[0092]

[0093] In equation (6), and It is the centroid of the data from two adjacent task phases in the i-th run. yes and The correlation coefficient, yes and covariance, and They are and The variance, v = 1, 2, ..., V i .

[0094] S32: Determine the correlation coefficient between the centroids of data from two adjacent task phases. Does it meet the relevance requirements? If satisfied, the adjustable parameter μ is updated according to μ = μ + 0.1, and formula (5) is returned. The initial value of the adjustable parameter μ is 4. Otherwise, the precisely divided task stages are obtained, and the running data is precisely divided to obtain C. i Each task phase, C i The data for each task phase is denoted as

[0095] In a specific embodiment, taking the data from the first run in this embodiment as an example, such as... Figure 6 As shown, it can be seen that as the adjustable parameter μ increases, the information increment δ of the searched mutations increases exponentially. hThe number of tasks will decrease, resulting in fewer task stages. When the initial value of the adjustable parameter μ is 4, the data from the first run is divided into 15 task stages, and the data from these 15 task stages is denoted as... Due to the task phase data in task phase 15 and Correlation coefficient between centroids The value is 0.97, which satisfies the requirement. Therefore, the adjustable parameter μ is updated according to μ = μ + 0.1. When the adjustable parameter μ is updated to 4.2, the first run is divided into 9 task phases, denoted as... The correlation coefficients between the centroids of data from any two adjacent task phases in each of the nine task phases were successively determined and found to be unsatisfactory. v = 1, 2, ..., 9, resulting in precisely divided task stages, denoted as...

[0096] S4: Through the window data matrix acquisition module, task phase division module, and task phase optimization module, the historical operational data of the aero-engine is precisely divided into each operational phase to obtain the task phase data from all operational data, such as... Figure 7 As shown, the task stages of each run are of unequal length, and the trend of task stage changes varies in each run.

[0097] In a specific embodiment, the data from the four runs are precisely divided. The data from the first run is divided into nine task stages, and the data from these nine task stages are denoted as follows: The second run yielded data for eight task phases, and the data for these eight task phases are denoted as follows: The data from the third run yielded 8 task phases, and the data for these 8 task phases are denoted as follows: The data from the fourth run yielded 10 task phases, and the data for these 10 task phases is denoted as follows:

[0098] S5: The task phase rearrangement module calculates the similarity of task phase data in all running data, rearranges the task phases with high similarity in different running data into one task phase, and determines each task phase after rearrangement.

[0099] In this embodiment, the task phase rescheduling includes the following steps:

[0100] S51: Arrange the data from the four runs in a top-down order with the same variable dimensions to obtain a matrix. Calculate matrix The mean and standard deviation of the data, where the mean of all data for the j-th variable is... The calculation formula is The standard deviation s of all data for the j-th variable j The calculation formula is k t =1,2,…,31050.

[0101] S52: According to the matrix mean and standard deviation s j Standardize the data for each task phase in all runtime data, where the data for the c-th task phase of the i-th runtime data is... The data for the standardized task phase is obtained through standardization. k c The j-th variable data of each sample point The standardized calculation formula is as follows:

[0102]

[0103] In equation (7), c = 1, 2, ..., C i , It is the number of sample points in the c-th task stage of the i-th run, j = 1, 2, ..., 14.

[0104] S53: Calculate any two standardized task phase data and The Euclidean distance between them is used to obtain the Euclidean distance matrix D. ic, (P×Q), i=1,2,…,4, c=1,2,…,C i , n=1,2,…,4, t=1,2,…,C n t≠c, n≠i, P and Q are respectively and The number of sample points, of which data any sample point and data any sample point Euclidean distance between them p,q The calculation formula is as follows:

[0105]

[0106] In equation (8), p = 1, 2, ..., P, q = 1, 2, ..., Q.

[0107] According to the Euclidean distance matrix D ic,nt (P×Q) defines the similarity S. ic,nt The formula is as follows:

[0108]

[0109] In equation (9), D ic,nt (p,q) represents the Euclidean distance matrix D. ic,nt The element in the p-th row and q-th column of S ic,nt The smaller the value, the higher the similarity between the two task phases, indicating that the two task phases perform the same task.

[0110] S54: For all running data, the standardized task stage data are similar to each other according to formulas (8) and (9). The data of the task stages with high similarity are expanded as variables and rearranged into a task stage, resulting in 8 task stages. The data of the 8 task stages are denoted as {X1,…,X...} sp ,…,X8}, where X sp This is the data for the sp-th task phase. K sp It is the number of sample points in the sp-th task stage, where sp = 1, 2, ..., 8.

[0111] S6: The model building module constructs a PCA monitoring sub-model for each of the eight task stages, and calculates the monitoring statistics and control limits for each task stage.

[0112] In this embodiment, the PCA method is used to decompose the data of each of the eight task stages, where the data X of the sp-th task stage is decomposed. sp The calculation formula for decomposition is as follows:

[0113]

[0114] In equation (10), For the data X of the sp-th task phase sp The estimate, W sp For the data X of the sp-th task phase sp The residual matrix, T so and P sp The data X for the sp-th task stage are respectively sp The score matrix and load matrix, sp = 1, 2, ..., 8.

[0115] Based on the data X of the sp-th task phase sp The score matrix T sp and load matrix P sp Calculate the monitoring statistic SPE for the sp-th task phase. sp The calculation formula is as follows:

[0116]

[0117] In equation (11), sp = 1, 2, ..., 8.

[0118] Based on all monitoring statistics SPE of the sp-th task phase sp Calculate the control limit SPE sp,α The formula is as follows:

[0119]

[0120] In equation (12), a sp The monitoring statistics SPE for the sp-th task phase. sp The mean, b sp The monitoring statistics SPE for the sp-th task phase. sp The variance.

[0121] S7: The online monitoring module determines whether the currently running online data is in a fault state based on the constructed PCA monitoring sub-model.

[0122] In this embodiment, determining whether the currently running online data is in a fault state involves the following steps:

[0123] S71: Obtain the current running data of 14 process variables for the l-th sample point of the new running process online. Use the mean and standard deviation of formula (7) to analyze the current running data. Standardization is performed to obtain the standardized current running data x. new,k x new,k =(x new,k,1 ,…,x new,k,14 ), where the j-th variable data of the k-th sample point. The standardized formula is as follows:

[0124]

[0125] In equation (13), and s j For matrix The mean and standard deviation of all data for the j-th variable, x new,k,j Let j be the standardized data of the j-th variable of the k-th sample point, where j = 1, 2, ..., 14.

[0126] S72: Use formulas (8) and (9) to determine the standardized current running data x new,k In the relevant task phase, calculate the standardized current running data x. new,k The Euclidean distance to the data at each task stage, where the standardized current running data x new,k Data X of the sp-th task stage sp any sample point Euclidean distance between The calculation formula is as follows:

[0127]

[0128] In equation (14), k sp =1,2,…,K sp sp = 1, 2, ..., 8, K sp It is the number of sample points in the sp-th task phase;

[0129] Calculate the standardized current running data x according to formula (14). new,k Data X of the sp-th task stage sp The Euclidean distance matrix D is obtained by calculating the Euclidean distance between each sample point in the matrix. new,ssp (1×K sp According to the Euclidean distance matrix D new,sp (1×K sp Define similarity S new,sp The formula is as follows:

[0130]

[0131] In equation (15), D new,sp (1,k sp ) represents the Euclidean distance matrix D new,sp The kth row of the first line sp Column elements.

[0132] S73: Calculate the standardized current operational data x based on the PCA monitoring sub-model of the relevant task phase. new,k Monitoring statistics SPE new The formula is as follows:

[0133]

[0134] In equation (16), For the standardized current running data x new,k The estimate, P sp For the data X of the sp-th task phase sp The load matrix, t sp This is the score vector.

[0135] S74: Calculate the monitoring statistic SPE new Control Limits (SPEs) of the Task Phase sp,α Comparison, if the monitoring statistic SPE new Not exceeding the control limit SPE sp,α If the current data is normal, it indicates that the current data is in a faulty state; otherwise, it indicates that the current data is in a faulty state.

[0136] To verify the effectiveness of this method, it was compared with the traditional global PCA method. Normal data and fault data were used for monitoring. The fault data included two types of faults, occurring at sample points 500, 991-1000, 3001-3010, and 6000, respectively. Table 3 shows the monitoring results of the traditional global PCA method and this method on the fault data. It can be seen that the traditional global PCA method sometimes fails to detect faults. Table 4 shows the monitoring performance of the traditional global PCA method and this method under normal and fault data conditions. As shown in Table 4, the monitoring method of this invention effectively reduces the false alarm rate and false negative rate, significantly improves the accuracy of fault detection, and has strong practical application in studying problems with unequal task phase lengths and differences in the changing trends of the operation process.

[0137] Table 3. Monitoring results of fault data using the traditional global PCA method and this method.

[0138]

[0139]

[0140] Table 4 shows the monitoring performance of the traditional global PCA method and the proposed method in two scenarios.

[0141]

Claims

1. An information increment-based method for monitoring the state of an aeroengine, characterized in that, Comprise the following steps: S1: the window data matrix acquisition module uses the window to select the single operation data in the aero-engine historical operation data, and obtains the window data matrix; S2: the task stage division module uses the information increment method to calculate the information increment index of the window data matrix, and then determines the switching point of the task stage by identifying the information increment index of the mutation, to obtain the divided task stage; The task stage division has the following steps: S21: The window data matrix uses an information increment method to calculate an information increment index. First, the window data matrix is preprocessed to obtain The calculation formula is as follows:​ In formula (1), b is the mean vector of T ∈R w×1 , k = 1, 2, …, K i -w+1; The pre-processed window data matrix The covariance matrix R is calculated k The calculation formula is as follows: Reuse two adjacent covariance matrix R h and Rh +1 Calculate the information increment matrix D h , the formula is as follows: D h = R h+1 - R h , 1≤h≤K i - w (3) Finally, the information increment matrix D h The information increment index δ h is calculated as follows: In formula (4), D h (a, b) represents an information increment matrix D h of the element in the ath row and the bth column, 1≤h≤K i -w, a = 1, 2…, J, b = 1, 2…, J, J is the number of process variables; K i is the number of sample points of the ith run; S22: using the window to select the information increment index δ h The selection is performed, and the window selects the continuous L information increment indexes δ h The window updates the information increment indexes in the window by moving back one information increment index each time, and sequentially judges whether each information increment index δ h in the window satisfies the mutation information increment index, and the judgment formula is as follows: δ h ≥ mean(d) + μ · std(d), d ≤ h ≤ L + d - 1 (5) In formula (5), mean(d) and std(d) are the mean value and the standard deviation of all information increment indexes δ h in the dth window, respectively, μ is an adjustable parameter, and d = 1, 2, …, K t -w-L; when the information increment index δ h satisfies formula (5), the information increment index δ h is the information increment index of the mutation, and all information increment indexes of the mutations are identified according to formula (5); S23: taking all the window data matrix corresponding to the information increment index of the mutation the final sample point in the middle as a switching point, obtaining all switching points, and dividing the running data according to all switching points to obtain V i task stages, and the data of the V i task stages are recorded as S3: the task stage optimization module adjusts the adjustable parameter value by judging whether the correlation coefficient between the centroids of the adjacent two task stage data in the divided task stage meets the correlation requirement, to obtain the accurately divided task stage; The task stage accurate division has the following steps: S31: Calculate the V i The correlation coefficient between the mass centers of two adjacent task stages in the task stage is calculated according to the following formula: in formula (6), and is the barycenter of the adjacent two task stage data in the i-th running data, is and is the correlation coefficient of is and is the covariance of and are the variances of and , respectively, v = 1, 2, …, V i ; S32: judging the correlation coefficient between the centers of the two adjacent task phase data in turn whether the correlation requirement is met If yes, the parameter μ is updated as μ = μ + 0.1, and returns to formula (5); otherwise, the task phase is accurately divided, and the running data is accurately divided to obtain C i task phases, and the data of the C i task phases are recorded as S4: through the window data matrix acquisition module, the task stage division module and the task stage optimization module, the each operation data in the aero-engine historical operation data is accurately divided, and the data of the task stage in all operation data is obtained; S5: the task stage rearrangement module calculates the similarity of the data of the task stage in all operation data, rearranges the task stages with high similarity in different operation data as one task stage, and determines each task stage after rearrangement; The task stage rearrangement has the following steps: S51: Arrange the data from I runs from top to bottom with the same variable dimensions to obtain a matrix. Calculate matrix The mean and standard deviation of the data, where the mean of all data for the j-th variable is... The calculation formula is The standard deviation s of all data for the j-th variable j The calculation formula is i = 1, 2, ..., I, k t =1,2,…,K t J is the number of process variables, K t K is the total number of sample points in the I-run data. i It is the number of sample points in the i-th run; S52: According to the matrix mean and standard deviation s j Standardize the data for each task phase in all runtime data, where the data for the c-th task phase of the i-th runtime data is... The data for the standardized task phase is obtained through standardization. k c The j-th variable data of each sample point The standardized calculation formula is as follows: In formula (7), c = 1, 2, …, C i , is the number of sample points of the cth task stage of the ith running data, j = 1, 2, …, J. S53: Calculate the Euclidean distance between any two normalized task phase data and to obtain the Euclidean distance matrix D ic,nt (P×Q), i = 1, 2, …, I, c = 1, 2, …, C i , n = 1, 2, …, I, t = 1, 2, …, C n , t≠c, n≠i, P and Q are the number of sample points of and respectively, wherein any sample point in data and any sample point in data The calculation formula of the Euclidean distance dist p,q between them is as follows: In formula (8), p=1, 2, …, P, q=1, 2, …, Q; According to the Euclidean distance matrix D ic,nt (P x Q) defines the similarity S ic,nt The formula is as follows: In formula (9), D ic,nt (p, q) denotes the Euclidean distance matrix D ic,nt the element in the pth row and qth column of matrix D S54: Similarity calculation is performed on the standardized data of all task stages in the running data according to formula (8) and formula (9), and the data of the task stages with high similarity are unfolded in a variable manner and rearranged into one task stage, and finally C * task stages are obtained. S6: constructing PCA monitoring sub-models for each of the C * task stages by the model construction module, and calculating the monitoring statistics and control limits for each of the task stages. S7: the online monitoring module judges whether the online current operation data is in a fault state according to the constructed PCA monitoring submodel; The judgment of whether the online current operation data is in a fault state has the following steps: S71: the current operation data is obtained online, and the mean value and the standard deviation in formula (7) are used to standardize the current operation data to obtain the standardized current data; S72: formula (8), formula (9) is used to judge the task stage to which the standardized current operation data belongs; S73: the monitoring statistic of the standardized current operation data is calculated according to the PCA monitoring submodel of the task stage; S74: the monitoring statistic of the current operation data is compared with the control limit of the task stage, if the monitoring statistic of the current operation data does not exceed the control limit of the task stage, it is explained that the current operation data is normal; Otherwise, it is judged that the current operation data is in a fault state.

Images