Mode recognition method for turboprop engine under continuous variable working condition

Abstract

The invention provides a mode recognition method for a turboprop engine under a continuous variable working condition. The mode recognition method comprises four parts of data processing, transition state separation, time sequence data segmentation and steady state mode recognition. The method simultaneously comprises a transition state mode identification function and a steady state mode identification function, and complete flight segment data is taken as a research object, so that the method has a data batch processing capability: in transition state mode identification, DeepSVDD and a CRNN network are combined, so that the method has a time sequence data processing capability, and the transition state mode is better segmented; in the steady-state mode identification, the method takes the minimum segmentation cost as a time series data segmentation criterion, and carries out steady-state mode identification by combining a label propagation algorithm. Compared with other engine mode recognition methods, the method has the advantages that the transition state data can be separated, so that the steady state mode recognition is not affected, and particularly, the method for recognizing the working mode of the engine in the actual situation is effectively constructed for a large amount of transition state data generated by continuous working condition changes in external field work of the aero-engine.

Description

Technical Field

[0001] The invention belongs to the field of aviation turboprop engine mode operation identification, and solves the problem of mode identification when the working conditions of an aviation engine continuously change during field use by designing a method for separating a transition state and a steady state. Background Art

[0002] Aircraft engine pattern recognition technology has significant significance, especially in improving flight safety, maintenance efficiency and reducing operating costs. First, by real-time monitoring and analysis of engine operating data, the technology can detect potential faults or abnormal conditions at an early stage, thereby preventing sudden faults and accidents, and significantly improving flight safety. Secondly, pattern recognition technology makes fault diagnosis more accurate and timely, and can provide airlines with scientific maintenance decision-making basis and optimize maintenance plans. This not only reduces unplanned downtime, but also extends the service life of the engine and improves operational efficiency. In addition, the implementation of preventive and predictive maintenance strategies helps to reduce maintenance costs and resource consumption, thereby improving economic benefits. Overall, the application of aircraft engine pattern recognition technology has greatly promoted the intelligence and modernization of aircraft maintenance management, with far-reaching economic and social benefits.

[0003] Aircraft engine pattern recognition technology is a key component of modern aviation maintenance and safety assurance systems. It aims to identify and predict the working status and potential failures of engines by analyzing and processing a large amount of data generated during engine operation. This technology mainly relies on the widespread deployment of sensor networks, which can monitor multiple key parameters of the engine in real time, such as temperature, pressure, vibration, fuel flow, etc. The collected data is processed through data preprocessing steps such as noise filtering, signal normalization, and outlier detection to ensure the accuracy of the analysis. Next, feature extraction technology is used to extract features that can effectively characterize the health status of the engine from the preprocessed data. These features may include frequency domain and time domain signal features. Pattern recognition is the core link, and machine learning algorithms such as support vector machines (SVM), neural networks, random forests, etc. are usually applied to classify and pattern match features to identify different operating modes. This information is crucial for predicting potential failures and formulating preventive maintenance strategies, which can significantly improve the safety and reliability of aircraft and reduce unplanned downtime and maintenance costs. With the advancement of big data and artificial intelligence technologies, aircraft engine pattern recognition technology is moving towards a more intelligent, real-time and precise direction, providing the aviation industry with more powerful tools to cope with complex engineering challenges.

[0004] In the field of aircraft engine pattern recognition, the engine will undergo a variety of operating conditions in the actual working environment, which causes the data collected by the sensor to contain a large amount of transition state data. Traditional analysis methods often ignore the impact of actual transition state data, resulting in inaccurate recognition results. Summary of the invention

[0005] In order to solve the current problem of pattern recognition after steady-state data acquisition of aircraft engines in the field, the present invention proposes an aircraft engine pattern recognition method (DeepMiner) that separates transition state and steady state, which can effectively solve the problem that it is difficult to accurately identify patterns of aircraft engines under continuously changing operating conditions.

[0006] The method for separating the transition state and the steady state of aircraft engine pattern recognition (DeepMiner) includes four parts: data processing, transition state separation, time series data segmentation and steady state pattern recognition. This method will include the recognition functions of both the transition state mode and the steady state mode, and take the complete flight segment data as the research object to enable it to have the data batch processing capability: in the transition state pattern recognition, this method combines the unsupervised learning algorithm DeepSVDD with the CRNN network to enable it to have the ability to process time series data, so as to better segment the transition state mode; in the steady state pattern recognition, this method uses the minimum segmentation cost as the time series data segmentation criterion, and performs steady state pattern recognition by combining the label propagation algorithm. Compared with other engine pattern recognition methods, DeepMiner can separate the transition state data so that the steady state pattern recognition is not affected, especially for the continuous working condition changes in the field work of aircraft engines, which will generate a large amount of transition state data, and effectively construct a method for engine working mode recognition in actual situations.

[0007] The present invention proposes a pattern recognition method for a turboprop engine under continuously variable operating conditions, comprising the following steps:

[0008] Step 1: Process the data collected from the field aircraft engine and merge these time series data to obtain a multi-dimensional time series;

[0009] The method comprises a multidimensional time series X=[x 1 , x 2 , …, x n ], x i is the d-dimensional vector at the i-th moment. The multidimensional time series X can be divided into K segments according to different modes, that is, the data segment set And get the mode timing segmentation point in The pattern labels corresponding to each data segment constitute a label set ε = {e 1 , e 2 ,…,e K}, where ei∈{1, 2, …, S} represents the i-th segment w i The working mode to which it belongs, s is the total number of all working modes including the transition state.

[0010] Necessary notes for step 1: The aircraft engine training set includes parameters: absolute pressure altitude H a , relative pressure altitude H r 、Total atmospheric temperature T t 、Atmospheric static temperature T s , Mach number M, speed Ng, true angle of attack θ, true airspeed V t , indicated airspeed V i , throttle angle PLA, torque τ, instantaneous consumption SFC, exhaust temperature EGT, and fuel pressure P.

[0011] Step 2, establish a DeepSVDD-CRNN transition state pattern recognition model;

[0012] Necessary notes for step 2: According to the standard steady-state operating mode definition in the engine manual, construct multi-dimensional steady-state mode time series data as the training data set in L is the time length of the sample, d is the feature dimension of the sample, and the data set Contains f p There are working modes, and the number of samples contained in each mode is equal.

[0013] Extract spatial features through convolutional layers

[0014]

[0015] Among them, Conv represents the convolution operation, W c Represents the weight of the convolutional layer. Next, the convolutional features are input into the recurrent layer to extract the temporal features.

[0016]

[0017] Among them, LSTM represents a recurrent neural network unit, W l represents the weight of the convolutional layer. Finally, The final feature representation is obtained through the function mapping parameterized by the neural network Will Substitute into the objective function of DeepSVDD:

[0018]

[0019] Where c is the center of the sphere, and the second term is L2 regularization. During the training process, the above objective function is minimized to make all samples closest to the center of the sphere to update the parameters of the DeepSVDD model. Calculate the distance between its feature representation f and the center c of the sphere:

[0020] s(f)=||φ(f)-c||

[0021] If it is greater than a certain threshold, it is judged that it does not belong to the same category as the training data set x, which is a transition state mode.

[0022] Step 3: Establish a time series segmentation algorithm based on minimum segmentation cost.

[0023] The method includes steady-state time series data Z=[Z 1 , Z 2 , …, Z p ] to perform time segment segmentation, where represents the ith point after rearrangement. Assume that Z can be divided into m segments by the inherent pattern of the data, that is, H = [h 1 ,h 2 ,…,h m ], where hh i is the split point λ i and λ i+1 The segmentation point set of sample Z can be expressed as Λ={λ 1 ,λ 2 , …, λ m+1}, and satisfy 1=λ 1 <λ 2 <…<λ m+1 =n. Each segment can be assigned a label according to its working mode, forming ={c 1 , c 2 , …, c m}, where c j ∈{1, 2, ..., f} represents the jth segment h j The working mode to which it belongs, and f is the total number of working modes.

[0024] Necessary notes for step 3: A mathematical model is built for each segment, and the models in different working modes will be significantly different. i The model can be expressed as:

[0025]

[0026] Among them, z t represents the segment h at time t in the time series Z i Belongs to working mode c i The sample, deviation b iThe mean is 0 and the variance is a d-order square matrix ∑ i The normal distribution of q is the order of the model. The present invention selects a first-order model, i.e., q=1, t i =[1, t-λ i ] T At this time, use As h i The parameter set of the model to be requested.

[0027] A(p, m) is used to represent a method of dividing p ordered steady-state data samples into m segments, and D is defined as M (i, j) represents the difference degree of data points in the same mode, that is, the loss. In order to find the optimal segmentation method, the following loss function is defined:

[0028]

[0029] Where i and j represent the starting data point and the ending data point in the same mode respectively; H represents the set of m segments of the steady-state data set, that is, H = [h 1 ,h 2 ,…,h m ],Z=[Z 1 , Z 2 , …, Z p ] represents a multidimensional steady-state data set, and P(Z|H) is the likelihood function value.

[0030] Considering the similarity of physical laws under the same working mode, according to the assumption of the deviation term, lnP(Z|H) can be converted to:

[0031]

[0032] Among them, |∑i| represents the determinant of the covariance matrix. Derivative the formula to obtain the extreme point, and the final result is:

[0033]

[0034] According to the characteristics of the acquired data set, the minimum number of steady-state data points is set to o. The optimal segmentation method is recorded as B(p, m). For the optimal segmentation L[B(p, m)], assuming that the last segment, i.e. the mth segment of the steady-state mode, contains only u data points u>o, the loss function of the optimal segmentation can be expressed as:

[0035] L[B(p,m)]=min m≤x≤p L[B(pu,m-1)]+D(x,p)

[0036] Among them, min m≤x≤pL[B(pu,m-1)] represents the optimal segmentation of the first (pu) data points into (m-1) classes, and D(x,p) represents the mth steady-state mode Z m =[z x , …, z p ] and p-x+1=u, and iteratively calculate the optimal segmentation point to minimize the segmentation loss.

[0037] Step 4: Establish a steady-state pattern recognition method based on label propagation.

[0038] Necessary comments for step 4: split the data segment h i Establish a linear model, model parameter set And the unlabeled sample points are composed of According to the engine steady-state mode judgment rules, create labeled multi-dimensional steady-state mode data as a label propagation reference data matrix in represents the eigenvector of the ith steady-state mode and has the same dimension as Φ i Same. At the same time, the label matrix corresponding to the reference data set is established can be regarded as a dimension f p The identity matrix, column vector The element in the i-th row is 1, which means that its corresponding steady-state working mode is i.

[0039] The unlabeled sample points With reference data matrix merge And according to the label propagation principle, an undirected graph is established. The nodes of the graph represent the similarity between the unlabeled data points and the reference data points. The labels can be propagated through the edges of the nodes. The larger the weight of the edge, the more similar the two nodes are, and the easier it is for the labels to propagate. and The propagation weight between jk , the formula is as follows:

[0040]

[0041] Where j≠k, α is a hyperparameter. According to this formula, we can further get the probability transfer calculation formula:

[0042]

[0043] Among them, P jk represents the transmission probability from node j to node k.

[0044] Set the label vector corresponding to the initial unlabeled data point to is a dimension fp The all-zero column vector corresponding to The label matrix To perform propagation calculation, the iterative formula is as follows:

[0045]

[0046] in Indicates the result of the nth iteration, reset The label matrix corresponding to the reference dataset is Repeat to get the result.

[0047] The propagation matrix P is divided as follows:

[0048]

[0049] Where P LU represents the probability of label propagation from reference data to unlabeled data points, and so on. The label after n iterations is represented as but The available formula is:

[0050]

[0051] After iterations, it converges to a convex solution:

[0052]

[0053] The obtained labels are soft labels, each item is the probability of the corresponding steady-state mode, and the row number corresponding to the maximum probability is selected as the sample point Belonging mode.

[0054] Beneficial effects: DeepMiner is mainly composed of three algorithms. The recognition method of transition state mode uses DeepSVDD-CRNN to model steady-state data to separate transition state data; the time series segmentation algorithm is based on the minimum segmentation cost. By modeling the data as a whole and iterating continuously to find the time series segmentation point with the least loss, the time series segments of different working modes are obtained; combined with the label propagation theory, the features of steady-state data of different working modes in the sample data set are extracted and the probability propagation matrix is ​​established to realize the recognition of different steady-state modes. The DeepMiner model is used in the task of aircraft engine pattern recognition to complete the recognition of different working states during the field use of aircraft engines. The pattern recognition experiment on the field use data of a certain type of turboprop engine shows that compared with the traditional data-driven method and many numerical calculation methods, the DeepMiner model is a more practical and effective method in the field of aircraft engine field pattern recognition. Considering the influence of various noise and transition state factors, it can avoid the inaccuracy caused by continuous working condition changes.

[0055] The DeepMiner method adopts a more practical and reliable method, that is, providing a transition state pattern recognition algorithm. This method can not only truly restore the nine working modes of the turboprop engine including the transition state, but also has an important impact on the subsequent real-time status monitoring, performance analysis and fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is the trend curve of the original data of a certain type of turboprop engine

[0057] Figure 2 It is the visual diagram of the working mode of segment a;

[0058] Figure 3 It is the visual diagram of the working mode of segment b;

[0059] Figure 4 It is the DeepMiner pattern recognition flow chart;

[0060] Figure 5 It is the flow chart of DeepSVDD-CRNN algorithm. DETAILED DESCRIPTION

[0061] This case takes a certain type of airborne turboprop engine as the research object. This type of engine is widely used in large transport aircraft and some rescue aircraft. Specifically, this type of engine is a fixed turbine single-rotor turboprop engine, which uses an axial flow compressor to provide compressed air for the combustion chamber and transmits the shaft power output by the engine to the propeller through a reducer. The flight data of 4 turboprop engines of a certain type during the entire field use cycle were counted. The flight parameter variables contained are shown in Table 1. PLA, Ng, SFC and EGT are selected as variables for engine pattern recognition. Figure 1 The data used in this case contains nine working modes, and the specific corresponding serial numbers are shown in Table 2, where the transition state data is regarded as one working mode, and the steady state data contains 8 engine working modes.

[0062] In order to verify the superiority of the proposed DeepMiner model in aircraft engine pattern mining, four algorithms were selected for comparative experiments:

[0063] 1) K-nearest neighbor algorithm (KNN): KNN algorithm is a commonly used pattern recognition algorithm, which performs pattern recognition by using distance measurement between samples.

[0064] 2) Support Vector Machine Algorithm (SVM): The SVM algorithm performs pattern recognition by finding the optimal hyperplane in the feature space.

[0065] 3) Gaussian Mixture Model (GMM): GMM assumes that the data is a mixture of multiple Gaussian distributions, each of which corresponds to a cluster or component of the data, for pattern recognition.

[0066] 4) Time Series Segmentation Algorithm (PELT-SOM): PELT-SOM is trained and used for recognition by mapping sample datasets into high-dimensional space and clustering them.

[0067] Table 1 Variables collected by sensors of a certain type of turboprop engine

[0068]

[0069] Table 2 Description of the working mode of a certain type of turboprop engine

[0070]

[0071]

[0072] This case uses four sets of turboprop engine data to conduct a comparative experiment. In the experiment, the four variables PLA, Ng, SFC and EGT are plotted on the same graph through dimensionless operation. The red dotted line is drawn according to the engineering manual to divide different working modes. The recognition results of different algorithms in segment a and segment b are plotted and compared as follows Figure 2 , Figure 3 As shown, from Figure 2 It can be seen that the KNN and SVM algorithms only learn sample data sets and do not consider transition state patterns. The GMM algorithm directly clusters the data set to obtain pattern recognition results, and the results are poor. The PELT-SOM algorithm has a strong recognition ability in data sets with a small number of patterns. The DeepMiner algorithm can almost completely recognize transition state data and the pattern segmentation is relatively accurate. Figure 3 It can be seen that when there are a large number of patterns and the data fluctuates greatly, the PELT-SOM algorithm will generate unnecessary data segmentation points during recognition, which also reduces its recognition ability. Because DeepMiner can identify the transition state in advance before the steady-state data is segmented in time series, its pattern recognition ability will not be affected too much, and its accuracy is still ahead of the other four algorithms.

[0073] Different algorithms are used to perform pattern recognition on the data of four engines. In order to fully demonstrate the effectiveness of the algorithms, this case selects accuracy, precision, recall and F1 score as evaluation indicators. The experimental results are averaged and compared, as shown in Table 3.

[0074] Table 3 Comparison of evaluation indicators of different algorithms

[0075]

[0076]

[0077] In Table 3, the accuracy, precision and recall are calculated by confusion matrix, and the average value is obtained by 10 experiments. The values ​​in brackets represent the mean square error of the corresponding evaluation index in the experiment. As can be seen from Table 3, the accuracy of DeepMiner algorithm and PELT-SOM algorithm that can identify transition state is much higher than that of other algorithms, among which DeepMiner is better than PELT-SOM in all evaluation indicators. When the aircraft engine performs flight missions, the driver's operation and the rapid change of external working conditions will affect the current working mode, which also leads to a large number of transition states between different steady-state modes. PELT-SOM only learns the distribution law of different mode data from the sample data set, but the sample data set does not contain transition state data, so its learned features are not enough to enable the model to fully identify the transition state mode, resulting in the inability to accurately cluster and judge the mode type in the case of multi-mode and large data noise. DeepMiner adopts a step-by-step recognition mode. By first identifying the transition state data and separating it, the steady-state mode recognition process is less disturbed and the accuracy is improved.

Claims

1. A pattern recognition method for a turboprop engine under continuously changing operating conditions, characterized in that: The steps include: Step 1: Processing the time series data of aircraft engine parameters collected from the field aircraft engine, and merging the time series data to obtain a multi-dimensional time series; Step 2, establishing a transition state mode recognition model, and according to the standard steady-state working mode definition in the engine work manual, constructing multidimensional steady-state mode time series data as a sample data set of the transition state mode recognition model, the transition state mode recognition model includes CRNN and DeepSVDD, CRNN is used to extract the time series features of the multidimensional steady-state mode time series data, and obtain the final feature representation through neural network parameterized function mapping, DeepSVDD uses the feature representation as a positive sample for model training; Step 3, using the transition state pattern recognition model trained in step 2 to process the multidimensional time series in step 1, and distinguishing the steady state pattern time series data from the transition state pattern time series data; Step 4, segmenting the steady-state mode time series data in step 3 into time series segments based on a time series segmentation algorithm with a minimum segmentation cost, establishing a linear model for each segmented data segment, and combining the parameter set of the linear model into unlabeled data points; Step 5, based on the multidimensional steady-state pattern time series data in step 2, create labeled multidimensional steady-state pattern data as a label propagation reference data matrix; based on the label propagation algorithm, calculate the similarity probability between each unlabeled data point and the reference data point in the label propagation reference data matrix, and the steady-state pattern of the reference data point corresponding to the maximum probability is the pattern to which the unlabeled data point belongs.

2. A pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 1, characterized in that: In step 1, the aircraft engine parameters include: absolute pressure altitude H a , relative pressure altitude H r 、Total atmospheric temperature T t 、Atmospheric static temperature T s , Mach number M, speed Ng, true angle of attack θ, true airspeed V t , indicated airspeed V i , throttle angle PLA, torque τ, instantaneous consumption SFC, exhaust temperature EGT, and fuel pressure P.

3. A pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 1, characterized in that: The step of training the transition state pattern recognition model in step 2 comprises: Construct multi-dimensional steady-state mode time series data as training data set in L is the time length of the sample, d is the feature dimension of the sample, and the data set x contains f p working modes, and the number of samples contained in each mode is equal; Extract spatial features through convolutional layers Among them, Conv represents the convolution operation, W c represents the weight of the convolutional layer; Input the convolutional features into the recurrent layer to extract the temporal features Among them, LSTM represents a recurrent neural network unit, H t represents the hidden state at time step t, W l represents the weight of the convolutional layer; Will The final feature representation is obtained through the function mapping parameterized by the neural network Will Substitute into the objective function of DeepSVDD: Where c is the center of the sphere, the second term is L2 regularization, and λ is a hyperparameter. During the training process, the above objective function is minimized to make all samples closest to the center of the sphere, so as to update the parameters of the entire model. For newly input test data Calculate the distance between its feature representation f and the center c of the sphere: s(f)=||φ(f)-c|| If it is greater than a certain threshold, it is judged that it does not belong to the same category as the training data set x, which is a transition state mode.

4. A pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 1, characterized in that: In step 4, the i-th segment h i The linear model is expressed as: Among them, z t represents the segment h at time t in the time series Z i Belongs to working mode c i The sample, deviation b i The mean is 0 and the variance is a d-order square matrix ∑ i The normal distribution of q is the order of the model; As h i The parameter set of the model to be requested.

5. A pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 4, characterized in that: The linear model is a first-order model, that is, q=1, t i =[1, t-λ i ] T ,λ i Represents sample z t The corresponding split point.

6. A pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 1, characterized in that: The specific steps of segmenting the steady-state mode time series data into time series segments in step 4 include: A(p, m) is used to represent a method of dividing p ordered steady-state data samples into m segments, and D is defined as M (i, j) represents the difference degree of data points in the same mode, that is, the loss. In order to find the optimal segmentation method, the following loss function is defined: Where i and j represent the starting data point and the ending data point in the same mode respectively; H represents the set of m segments into which the steady-state data set is divided, i.e., H = [h1, h2, …, h m ],Z=[Z1,Z2,…,Z p ] represents a multidimensional steady-state data set, P(Z|H) is the likelihood function value; Considering the similarity of physical laws under the same working mode, according to the assumption of the deviation term, lnP(Z|H) is converted into: Among them, |∑ i | represents the determinant of the covariance matrix; taking the derivative of the formula to obtain the extreme point, we get:

7. A pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 6, characterized in that: According to the characteristics of the acquired data set, the minimum number of steady-state data points is set to o; the optimal segmentation method is recorded as B(p, m). For the optimal segmentation L[B(p, m)], assuming that the last segment, i.e. the mth segment of the steady-state mode, contains only u data points u>o, the loss function of the optimal segmentation is expressed as: L[B(p,m)]=min m≤x≤p L[B(p-u,m-1)]+D(x,p) Among them, min m≤x≤p L[B(pu,m-1)] represents the optimal segmentation of the first (pu) data points into (m-1) classes, and D(x,p) represents the mth steady-state mode Z m =[z x , …, z p ] and p-x+1=u; Iterate and calculate the optimal segmentation point to minimize the segmentation loss. The data divided by the optimal segmentation point is the data segment of different steady-state modes, and each segment corresponds to a different Φ i for its parameter set.

8. The pattern recognition method for a turboprop engine under continuously variable operating conditions according to claim 1, characterized in that: Step 5 specifically includes: According to the engine steady-state mode judgment rules, create labeled multi-dimensional steady-state mode data as a label propagation reference data matrix in represents the eigenvector of the ith steady-state mode and has the same dimension as Φ i Same; at the same time, establish the label matrix corresponding to the reference data set is a dimension f p The identity matrix, column vector The element in the i-th row is 1, which means that its corresponding steady-state working mode is i; The unlabeled sample points With reference data matrix merge An undirected graph is established based on the label propagation principle. The nodes of the graph represent the similarity between the unlabeled data points and the reference data points. The labels are propagated through the edges of the nodes. The larger the weight of the edge, the more similar the two nodes are, and the easier it is for the labels to propagate. and The propagation weight between jk , the formula is as follows: Where j≠k, α is a hyperparameter, and the probability transfer calculation formula is obtained according to this formula: Among them, P jk represents the transmission probability from node j to node k; Set the label vector corresponding to the initial unlabeled data point to is a dimension f p The all-zero column vector corresponding to The label matrix To perform propagation calculation, the iterative formula is as follows: in Indicates the result of the nth iteration, reset The label matrix corresponding to the reference dataset is part, repeat to get the result; The propagation matrix P is divided as follows: Where P LU represents the probability of label propagation from reference data to unlabeled data points, and so on; the label after n iterations is represented as but The formula is: After iterations, it converges to a convex solution: The obtained labels are soft labels, each item is the probability of the corresponding steady-state mode, and the row number corresponding to the maximum probability is selected as the sample point The mode.

Images