An air path fault hypersphere clustering method based on angle interval constraint

Abstract

The application belongs to the technical field of aero-engine gas path fault diagnosis, and discloses a gas path fault hyperspherical clustering method based on angle interval constraint, which comprises the following steps: S1, obtaining gas path fault data samples and fault-free data samples; S2, extracting features in the gas path fault data samples and the fault-free data samples to obtain sample features; S3, calculating feature residuals according to the sample features, and calculating residual differences according to the feature residuals; S4, constructing an angle interval constraint loss function, an L2 norm constraint loss function and a class center aggregation function; S5, constructing a total loss function; and S6, inputting the sample features, the feature residuals and the residual differences into the total loss function to obtain a hyperspherical clustering result. The application improves the accuracy and robustness of gas path fault diagnosis and provides assistance for engine health management and gas path fault analysis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of aero-engine gas path fault diagnosis, and particularly relates to a gas path fault hyperspherical clustering method based on angle interval constraints. BACKGROUND

[0002] An aero-engine is a whole composed of complex systems, and in the full envelope range thereof, a single parameter is difficult to reflect the advantages and disadvantages of the engine gas path characteristics and performance, and it is necessary to analyze the ratio between different cross-section parameters, and the performance dispersion degree between engines is also large, so that the cross-section parameter change rule caused by each gas path fault is not obvious, and it is difficult to accurately distinguish the fault characteristics by using conventional methods. SUMMARY

[0003] To solve the above technical problems, the application provides a gas path fault hyperspherical clustering method based on angle interval constraints, comprising the following steps:

[0004] S1, obtaining gas path fault data samples and fault-free data samples;

[0005] S2, extracting features in the gas path fault data samples and the fault-free data samples to obtain sample features;

[0006] The sample features include a plurality of gas path fault features and fault-free features.

[0007] S3, calculating feature residuals according to the sample features, wherein the feature residuals include gas path fault feature residuals and fault-free feature residuals, and calculating residual differences according to the feature residuals;

[0008] S4, constructing an angle interval constraint loss function, an L2 norm constraint loss function, and a class center aggregation function;

[0009] S5, constructing a total loss function according to the angle interval constraint loss function, the L2 norm constraint loss function, and the class center aggregation function;

[0010] S6, inputting the sample features, the feature residuals, and the residual differences into the total loss function to obtain a hyperspherical clustering result, wherein the hyperspherical clustering result includes a fault-free feature cluster and a plurality of residual difference clusters.

[0011] Preferably, the gas path fault data samples include parameters measured by sensors in a gas path fault state and reference model output parameters.

[0012] The parameters measured by the sensors in the fault state of the gas path include: fault engine inlet total temperature, fault fan rotating speed, fault compressor rotating speed, fault fan inlet static pressure, fault compressor rear static pressure, fault low-pressure turbine rear pressure, fault low-pressure turbine rear total temperature, fault main fuel flow, fault afterburner fuel flow and fault nozzle area;

[0013] The fault-free data samples include: parameters measured by the sensors in the fault-free state and reference model output parameters;

[0014] The parameters measured by the sensors in the fault-free state include: fault-free engine inlet total temperature, fault-free fan rotating speed, fault-free compressor rotating speed, fault-free fan inlet static pressure, fault-free compressor rear static pressure, fault-free low-pressure turbine rear pressure, fault-free low-pressure turbine rear total temperature, fault-free main fuel flow, fault-free afterburner fuel flow and fault-free nozzle area;

[0015] The reference model output parameters include: standard engine inlet total temperature, standard fan rotating speed, standard compressor rotating speed, standard fan inlet static pressure, standard compressor rear static pressure, standard low-pressure turbine rear pressure, standard low-pressure turbine rear total temperature, standard main fuel flow, standard afterburner fuel flow and standard nozzle area.

[0016] Preferably, the S2, the features in the gas path fault data samples and the fault-free data samples, and the acquisition of the sample features include the following steps:

[0017] S21, extracting the features in the gas path fault data samples and the fault-free data samples, and acquiring high-dimensional data features;

[0018] S22, reducing the high-dimensional data features to three-dimensional feature vectors, and acquiring sample features.

[0019] Preferably, the gas path fault features are composed of the correlation between the gas path fault data sample parameters, including: gas path fault sensor measurement features and reference model output features;

[0020] The fault-free features are composed of the correlation between the fault-free data sample parameters, including: fault-free sensor measurement features and reference model output features.

[0021] Preferably, the calculation of the feature residual according to the sample features includes the following steps:

[0022] S31, calculating the gas path fault feature residual according to the gas path fault features;

[0023] The gas path fault feature residual specifically refers to the difference between the gas path fault sensor measurement features and the reference model output features;

[0024] S32, calculating the fault-free feature residual according to the fault-free features;

[0025] The fault-free feature residual specifically refers to a difference between a fault-free sensor measurement feature and a reference model output feature.

[0026] The residual difference specifically refers to a difference between the gas path fault feature residual and the fault-free feature residual.

[0027] Preferably, the constructing the angle interval constraint loss function comprises the following steps:

[0028] S41, introducing a minimum angle interval parameter;

[0029] S42, setting a same-class sample feature angle threshold and a different-class sample angle threshold according to the minimum angle interval parameter;

[0030] S43, constructing an angle interval constraint loss function according to the same-class sample feature angle threshold and the different-class sample angle threshold.

[0031] Preferably, the angle interval constraint loss function is used to force the angle between same-class sample features to be less than the same-class sample feature angle threshold and force the angle between different-class samples to be greater than the different-class sample angle threshold.

[0032] The angle interval constraint loss function The expression is as follows:

[0033]

[0034]

[0035]

[0036] In the formula: is a loss value, is a sample feature, is a category label corresponding to the sample feature, is a minimum angle interval parameter, is a judgment condition, and i and j are sample serial numbers.

[0037] Preferably, the L2 norm constraint loss function expression is as follows:

[0038]

[0039] In the formula: is a number of fault-free sample features, is a number of nth fault sample features, C is a total number of fault types, and r is a hypersphere radius, is a fault-free feature residual, is a fault feature residual, is a residual difference.

[0040] Preferably, the class center aggregation function The expression is as follows:

[0041]

[0042] In the formula: Let C be the number of fault samples of type n, and C be the total number of fault types. The difference between residuals It is the center of the difference of residuals.

[0043] Preferably, the total loss function The expression is as follows:

[0044]

[0045] In the formula: The scaling factor of the L2 norm constrained loss function, L2 norm constrained loss function The proportional function of the class center aggregation function, For class center aggregation functions, The scaling factor of the angle interval constraint loss function. This is the loss function for the angular interval constraint.

[0046] The beneficial effects of this invention are as follows:

[0047] It can effectively extract the characteristics of gas path faults in aero-engines, cluster similar faults, separate dissimilar faults, and ensure that the center of each fault data is evenly distributed on a hypersphere. Geometric constraints enhance the interpretability of fault diagnosis tasks and improve the accuracy and robustness of aero-engine gas path fault diagnosis. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the system structure provided in an embodiment of the present invention;

[0049] Figure 2 The hyperspherical clustering results provided in the embodiments of the present invention are viewed in front of the present invention. Figure 1 ;

[0050] Figure 3 The hyperspherical clustering results provided in the embodiments of the present invention are viewed in front of the present invention. Figure 2 ;

[0051] Figure 4 This is a top view of the hyperspherical clustering results provided in an embodiment of the present invention;

[0052] Figure 5 This is an isometric projection of the hyperspherical clustering results provided in this embodiment of the invention. Detailed Implementation

[0053] The technical solutions of the present application are further described below, but the scope of protection is not limited to the description.

[0054] As shown in the figure, an air path fault hypersphere clustering method based on angle interval constraint comprises the following steps: Figure 1

[0055] S1, acquiring air path fault data samples and fault-free data samples;

[0056] The air path fault data samples comprise parameters measured by sensors in an air path fault state and reference model output parameters;

[0057] The parameters measured by the sensors in the air path fault state comprise fault engine inlet total temperature, fault fan rotational speed, fault compressor rotational speed, fault fan inlet static pressure, fault compressor rear static pressure, fault low-pressure turbine rear pressure, fault low-pressure turbine rear total temperature, fault main fuel flow, fault afterburner fuel flow, and fault nozzle area;

[0058] The fault-free data samples comprise parameters measured by sensors in a fault-free state and reference model output parameters;

[0059] The parameters measured by the sensors in the fault-free state comprise fault-free engine inlet total temperature, fault-free fan rotational speed, fault-free compressor rotational speed, fault-free fan inlet static pressure, fault-free compressor rear static pressure, fault-free low-pressure turbine rear pressure, fault-free low-pressure turbine rear total temperature, fault-free main fuel flow, fault-free afterburner fuel flow, and fault-free nozzle area;

[0060] The reference model output parameters comprise standard engine inlet total temperature, standard fan rotational speed, standard compressor rotational speed, standard fan inlet static pressure, standard compressor rear static pressure, standard low-pressure turbine rear pressure, standard low-pressure turbine rear total temperature, standard main fuel flow, standard afterburner fuel flow, and standard nozzle area.

[0061] In this embodiment, the sensor measurement parameters are denoted as , the reference model output parameters are denoted as . Further, the parameters measured by the sensors in the air path fault state are denoted as , the parameters measured by the sensors in the fault-free state are denoted as , and the reference model output parameters are denoted as .

[0062] S2, extracting features in the air path fault data samples and the fault-free data samples to obtain sample features;

[0063] The S2, extracting features in the air path fault data samples and the fault-free data samples to obtain sample features comprises the following steps:​

[0064] S21. Extract features from gas path fault data samples and fault-free data samples to obtain high-dimensional data features;

[0065] S22. Reduce the dimensionality of high-dimensional data features to three-dimensional feature vectors to obtain sample features.

[0066] In this embodiment, the high-dimensional data characteristics are reduced to three-dimensional feature vectors by using the t-distributed random neighborhood embedding method, also known as t-SNE.

[0067] The sample features include: several types of gas path fault features and fault-free features;

[0068] Gas circuit fault characteristics are divided into several categories, such as fault A, fault B, ..., fault N, based on gas circuit fault data samples.

[0069] In this embodiment, there are three types of gas path fault data samples, namely fault A, fault B and fault C.

[0070] The gas path fault characteristics are composed of the correlation between parameters of the gas path fault data samples, including: gas path fault sensor measurement characteristics and baseline model output characteristics;

[0071] The fault-free characteristics are composed of the correlation between parameters of fault-free data samples, including: fault-free sensor measurement characteristics and benchmark model output characteristics.

[0072] In this embodiment, the sensor measurement features are denoted as... The output features of the baseline model are denoted as Furthermore, the measurement characteristics of the gas path fault sensor are denoted as... The parameters measured by the sensor under fault-free conditions are recorded as follows: The output features of the baseline model are denoted as , where n is the fault type number.

[0073] S3. Calculate the feature residuals based on the sample characteristics. The feature residuals include gas path fault feature residuals and fault-free feature residuals. Calculate the difference between the residuals based on the feature residuals.

[0074] The calculation of feature residuals based on sample features includes the following steps:

[0075] S31. Calculate the residual of gas path fault characteristics based on the gas path fault characteristics;

[0076] The gas path fault characteristic residual specifically refers to the difference between the gas path fault sensor measurement characteristics and the baseline model output characteristics;

[0077] In this embodiment, the residual characteristic of the gas path fault is denoted as... wherein n is a fault type number.

[0078] S32, calculating a fault-free feature residual according to the fault-free feature;

[0079] In this embodiment, the fault-free feature residual is denoted as

[0080] The fault-free feature residual specifically refers to a difference between a fault-free sensor measurement feature and a reference model output feature.

[0081] The residual difference specifically refers to a difference between the air path fault feature residual and the fault-free feature residual.

[0082] The difference between the air path fault feature residual and the fault-free feature residual is the residual difference, denoted as .

[0083] S4, constructing an angle interval constraint loss function, an L2 norm constraint loss function, and a class center aggregation function;

[0084] The construction of the angle interval constraint loss function includes the following steps:

[0085] S41, introducing a minimum angle interval parameter;

[0086] In this embodiment, the minimum angle interval parameter is denoted as .

[0087] S42, setting a same-class sample feature angle threshold and a different-class sample angle threshold according to the minimum angle interval parameter;

[0088] The same-class sample feature angle threshold is , and the different-class sample angle threshold is . Wherein is an initial angle of the same-class sample feature, is an initial angle of the different-class sample.

[0089] S43, constructing the angle interval constraint loss function according to the same-class sample feature angle threshold and the different-class sample angle threshold.

[0090] The angle interval constraint loss function is used to force the angle between the same-class sample features to be less than the same-class sample feature angle threshold , and to force the angle between the different-class samples to be greater than the different-class sample angle threshold .

[0091] Thus, through the angle interval constraint loss function, the geodesic distance of the same-class sample features is minimized in the hyperspherical feature space, while the interval of the different-class sample features is maximized, thereby improving the intra-class compactness and inter-class separability.

[0092] The angle interval constraint loss function The expression is as follows:

[0093]

[0094]

[0095]

[0096] In the formula: is a loss value, is a sample feature, is a class label corresponding to the sample feature, is a minimum angle interval parameter, is a judgment condition, and i and j are sample serial numbers.

[0097] The expression of the L2 norm constraint loss function is as follows:

[0098]

[0099] In the formula: is a number of fault-free sample features, is a number of nth class fault samples, C is a total number of fault types, and r is a hyperspherical radius, is a fault-free feature residual, is a fault feature residual, is a difference between residuals.

[0100] The L2 norm constraint loss function is used to realize that the module length of the fault-free feature residual tends to 0, and the module length of the difference between all fault feature residuals and the fault-free feature residual tends to a constant value.

[0101] That is, the difference between residuals is located on a hypersphere with the fault-free feature residual as the center.

[0102] The class center aggregation function The expression is as follows:

[0103]

[0104] In the formula: is a number of nth class fault samples, C is a total number of fault types, is a difference between residuals, is a center of the difference between residuals.

[0105] wherein called class center, is a random number at initialization and gradually updates towards the center of the current batch of samples during the training process. .

[0106] The specific update formula is:

[0107]

[0108]

[0109] wherein: is the class center before iteration, is the class center after iteration, is the class center update rate, is the number of the nth class of fault samples.

[0110] The same class of samples is aggregated through the class center aggregation function, and the centers of different classes are evenly distributed on the hypersphere.

[0111] S5, constructing a total loss function according to the angle interval constraint loss function, the L2 norm constraint loss function and the class center aggregation function;

[0112] The total loss function is expressed as follows:

[0113]

[0114] wherein: is the proportion coefficient of the L2 norm constraint loss function, is the L2 norm constraint loss function, is the proportion function of the class center aggregation function, is the class center aggregation function, is the proportion coefficient of the angle interval constraint loss function, is the angle interval constraint loss function.

[0115] The total loss function is composed of the angle interval constraint loss function, the L2 norm constraint loss function and the class center aggregation function. Through the total loss function, the unordered three-dimensional feature vector, i.e. the sample feature, can be clustered to form a hypersphere with the fault-free sample feature as the center of the sphere, and the residual difference clusters are evenly distributed on the surface of the hypersphere. Finally, different types of faults are clustered and distinguished, and the effective extraction of fault features is realized.

[0116] S6, inputting the sample feature, the feature residual and the residual difference into the total loss function to obtain a hypersphere clustering result, wherein the hypersphere clustering result includes a fault-free feature cluster and a plurality of residual difference clusters.

[0117] As Figures 2 to 5 ​​The image shows the hyperspherical clustering results obtained in this embodiment. The difference in residuals is obtained by constraining the loss function using the L2 norm. Located in the residual with fault-free characteristics On a hypersphere with the center of the sphere, class aggregation functions are used to achieve class-based aggregation. Aggregation, Heterogeneous The centers of the faults are evenly distributed on the hypersphere. An angular interval constraint loss function is used to minimize the geodesic distance between features of the same class of samples in the hypersphere feature space, while maximizing the interval between features of different classes of samples, thereby improving intra-class compactness and inter-class separability. Finally, fault-free feature clusters are located at the center of the sphere, and residual difference clusters are distributed on the hypersphere surface, enabling the clustering and differentiation of various faults. When unknown gas path fault data enters, it can accurately classify the gas path fault data into the corresponding category.

[0118] This invention addresses complex aero-engine gas path fault diagnosis by employing an angular interval contrast loss function, along with an L2 norm constraint loss function and a class center aggregation loss function. By enhancing the interpretability of image classification tasks through geometric constraints, it minimizes the geodesic distance between gas path fault samples of the same type while maximizing the interval between samples of different types, thereby improving the compactness of samples of the same type and the classification accuracy between samples of different types. This method is then used to analyze the difference between fault feature residuals and fault-free feature residuals. The analysis was performed to determine the even distribution of different classes. On a hypersphere, and controlling the distance from the center of the sphere and the fault-free characteristic residual as the center of the sphere. The L2 norm approaches 0, which leads to the aggregation of similar fault samples and the uniform distribution of the centers of dissimilar fault samples on the hypersphere. This can solve the problem of difficulty in accurately extracting the characteristics of engine gas path faults, and lay the groundwork for accurate localization of aero-engine gas path faults. It improves the accuracy and robustness of gas path fault diagnosis and provides assistance for engine health management and gas path fault analysis.

Claims

1. A hyperspherical clustering method for gas path faults based on angular interval constraints, characterized in that, Includes the following steps: S1. Obtain gas path fault data samples and fault-free data samples; S2. Extract features from the gas path fault data samples and fault-free data samples to obtain sample features; The sample features include gas path fault features and fault-free features; S3. Calculate the feature residuals based on the sample characteristics. The feature residuals include gas path fault feature residuals and fault-free feature residuals. Calculate the difference between the residuals based on the feature residuals. S4. Construct the angle interval constraint loss function, the L2 norm constraint loss function, and the class center aggregation function; S5. Construct the total loss function based on the angle interval constraint loss function, the L2 norm constraint loss function, and the class center aggregation function; S6. Input the sample features, feature residuals, and residual differences into the total loss function to obtain the hyperspherical clustering results. The hyperspherical clustering results include fault-free feature clusters and residual difference clusters. The gas path fault data sample includes: parameters measured by the sensor under gas path fault conditions and baseline model output parameters; The parameters measured by the sensor under gas path failure conditions include: total temperature at the inlet of the faulty engine, speed of the faulty fan, speed of the faulty compressor, static pressure at the inlet of the faulty fan, static pressure after the faulty compressor, pressure after the faulty low-pressure turbine, total temperature after the faulty low-pressure turbine, main fuel flow rate, afterburner fuel flow rate, and nozzle area. The fault-free data sample includes: parameters measured by the sensor under fault-free conditions and output parameters of the benchmark model; The parameters measured by the sensor under fault-free conditions include: fault-free engine inlet total temperature, fault-free fan speed, fault-free compressor speed, fault-free fan inlet static pressure, fault-free compressor after static pressure, fault-free low-pressure turbine after pressure, fault-free low-pressure turbine after total temperature, fault-free main fuel flow rate, fault-free afterburner fuel flow rate, and fault-free nozzle area. The benchmark model output parameters include: standard engine inlet total temperature, standard fan speed, standard compressor speed, standard fan inlet static pressure, standard compressor post-static pressure, standard low-pressure turbine post-pressure, standard low-pressure turbine post-total temperature, standard main fuel flow rate, standard afterburner fuel flow rate, and standard nozzle area. The construction of the angle interval constraint loss function includes the following steps: S41, introducing the minimum angle interval parameter; S42. Set the feature angle threshold for similar samples and the angle threshold for dissimilar samples based on the minimum angle interval parameter; S43. Construct an angle interval constraint loss function based on the feature angle thresholds of similar samples and the angle thresholds of dissimilar samples; The angle interval constraint loss function is used to: force the angle between features of the same type of samples to be less than the angle threshold of features of the same type of samples, and force the angle between features of different types of samples to be greater than the angle threshold of features of different types of samples. The angle interval constraint loss function The expression is as follows: In the formula: For loss value, For sample features, For the category labels corresponding to the sample features, For the minimum angular interval parameter, The condition is denoted by i, and the sample numbers are i and j.

2. The hyperspherical clustering method for gas path faults based on angular interval constraints as described in claim 1, characterized in that, S2, extracting features from gas path fault data samples and fault-free data samples to obtain sample features includes the following steps: S21. Extract features from gas path fault data samples and fault-free data samples to obtain high-dimensional data features; S22. Reduce the dimensionality of high-dimensional data features to three-dimensional feature vectors to obtain sample features.

3. The hyperspherical clustering method for gas path faults based on angular interval constraints as described in claim 2, characterized in that, The gas path fault characteristics are composed of the correlation between parameters of the gas path fault data samples, including: gas path fault sensor measurement characteristics and baseline model output characteristics; The fault-free characteristics are composed of the correlation between parameters of fault-free data samples, including: fault-free sensor measurement characteristics and benchmark model output characteristics.

4. The hyperspherical clustering method for gas path faults based on angular interval constraints as described in claim 3, characterized in that, The calculation of feature residuals based on sample features includes the following steps: S31. Calculate the residual of gas path fault characteristics based on the gas path fault characteristics; The gas path fault characteristic residual specifically refers to the difference between the gas path fault sensor measurement characteristics and the baseline model output characteristics; S32. Calculate the fault-free characteristic residuals based on the fault-free characteristics; The fault-free feature residual specifically refers to the difference between the measurement features of the fault-free sensor and the output features of the benchmark model; The difference in residuals specifically refers to the difference between the residuals of gas path fault characteristics and the residuals of no-fault characteristics.

5. The hyperspherical clustering method for gas path faults based on angular interval constraints as described in claim 4, characterized in that, The L2 norm-constrained loss function is expressed as follows: In the formula: For the number of features of fault-free samples, Let C be the number of fault samples of type n, C be the total number of fault types, and r be the radius of the hypersphere. For fault-free characteristic residuals, For fault characteristic residuals, It is the difference of residuals.

6. The hyperspherical clustering method for gas path faults based on angular interval constraints as described in claim 5, characterized in that, The class center aggregation function The expression is as follows: In the formula: Let C be the number of fault samples of type n, and C be the total number of fault types. The difference between residuals It is the center of the difference of residuals.

7. The hyperspherical clustering method for gas path faults based on angular interval constraints as described in claim 6, characterized in that, The total loss function The expression is as follows: In the formula: The scaling factor of the L2 norm constrained loss function, L2 norm constrained loss function The proportional function of the class center aggregation function, For class center aggregation functions, The scaling factor of the angle interval constraint loss function. This is the loss function for the angular interval constraint.

Images