Multi-source data fusion axial flow compressor instability prediction method

Abstract

The invention discloses a multi-source data fusion axial flow compressor instability prediction method, and belongs to the technical field of aero-engine modeling and simulation. The method comprises the following steps: firstly, preprocessing dynamic pressure sensor data, constructing a multi-source data set and dividing the multi-source data set; secondly, respectively constructing a time domain feature extraction network module, a time-frequency domain image feature extraction network module and an interactive fusion module based on a cross-domain attention mechanism; thirdly, on the basis of the three modules, a multi-source data fusion compressor instability prediction model based on a cross-domain attention mechanism is constructed, and the model is trained to save an optimal model; and finally, performing real-time prediction on test data. According to the method, data of multiple sources are comprehensively processed and analyzed, complementarity and relevance among different data sources are captured, more comprehensive, complete and consistent perception about a target observation object is formed, diagnosis errors caused by a single data source are reduced, and the accuracy and reliability of a prediction result are improved.

Description

Technical Field

[0001] The present invention relates to a multi-source data fusion axial flow compressor instability prediction method, in particular to a multi-source data fusion axial flow compressor instability prediction method based on a cross-domain attention mechanism, and belongs to the technical field of aero-engine modeling and simulation. Background Art

[0002] As one of the three core components of an aircraft engine, the stable operation of the compressor is crucial to the engine performance and safety, and directly affects the overall performance of the engine. Compressor instability is a common gas path fault with extremely strong destructiveness and extremely fast changes. Therefore, its prediction has always been a research hotspot and difficulty in the field of international aircraft engines. Generally, the aerodynamic instability of a compressor mainly refers to two states: rotating stall and surge. When the gas flows into the compressor axially from the air inlet at a certain speed, the compressor compresses the gas through the compression component, thereby providing high-temperature and high-pressure gas to the combustion chamber. At this time, as the flow rate decreases and the total pressure ratio increases, it is very easy to cause gas flow separation, resulting in aerodynamic instability inside the compressor, and entering the stall precursor, rotating stall and surge state from the steady state. The above two aerodynamic instability states may cause the compressor performance to decline and the vibration to increase, which is an important factor restricting its reliable operation and may even damage the engine. This will not only increase the maintenance cost, but also may pose a serious threat to flight safety. In actual operation, we should try our best to avoid their occurrence. Therefore, in order to reserve more time for active control, it is of great significance to detect and identify compressor instability.

[0003] In order to monitor the operating status of multiple components of the compressor, multiple independent sensors are often installed inside it. However, traditional compressor fault diagnosis methods are usually based on a single data source, such as dynamic pressure signals, vibration sensor data, exhaust composition analysis data, etc. For example, the invention patent with publication number CN113836817A discloses a method for predicting rotating stall of an axial flow compressor based on a stacked long short-term memory network, using a single time-domain surge data to predict stall through a stacked long short-term memory network; the invention patent with publication number CN112001128B discloses a method for predicting stall and surge of an axial flow compressor based on deep learning, using dynamic pressure data and its time-domain statistical characteristics, and realizing surge prediction by constructing a prediction model based on dilated causal convolution and logistic regression. These methods often fail to fully reflect the complexity and variability of the engine's operating state. At the same time, these methods also ignore the correlation and redundancy between data collected by different sensors and the interaction between cross-domain information, and fail to mine the intrinsic connection between multi-source data. Moreover, stall warning involves complex minor fault identification, and the mechanism of the initial stall disturbance is very complex. Data from a single sensor may not be able to fully reflect changes in compressor operating conditions, and there are deficiencies in information completeness, accuracy, and timeliness. Summary of the invention

[0004] In view of the problems of insufficient accuracy, poor completeness and insufficient information exchange of single sensor data in the prior art, the present invention provides a multi-source data fusion axial flow compressor instability prediction method. The multi-source data fusion axial flow compressor instability prediction method provided by the present invention can mine the potential correlation of cross-domain information, enhance information completeness, and improve the stall surge prediction accuracy.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A multi-source data fusion axial flow compressor instability prediction method, specifically a multi-source data fusion axial flow compressor instability prediction method based on a cross-domain attention mechanism, comprising the following steps:

[0007] Step S1, preprocessing the dynamic pressure sensor data in the compressor gasping experiment of a certain type of aircraft engine, dividing it into two types of data, including time domain data and time-frequency domain image data, the multi-source data in the present invention refers to the time domain data and the time-frequency domain image data, including the following steps:

[0008] Step S1.1, obtaining dynamic pressure sensor data in the compressor gasping experiment, including data in a stable state, an initial disturbance state, and an unstable state, and normalizing, filtering, and downsampling the dynamic pressure sensor data to obtain time domain data;

[0009] Step S1.2, according to the sliding segmentation method, with the time step size as The time domain data obtained in step S1.1 is segmented, and each sample after segmentation is labeled with 1 or 0 according to whether there is surge or not, wherein the label with surge is 1, and the label without surge is 0;

[0010] Step S1.3, using the wavelet analysis method to generate the time-frequency domain image data of the corresponding time step, and labeling each sample after segmentation with 1 or 0 according to whether there is surge or not, wherein the label with surge is 1, and the label without surge is 0;

[0011] Step S1.4, the multi-source data set consists of time domain data and time-frequency domain image data in pairs. The data set is divided into a training data set and a test data set at a ratio of 4 to 5:1, and then the training data set is divided into a training set and a validation set at a ratio of 4 to 5:1.

[0012] Step S2, constructing a time domain feature extraction network module, includes the following steps:

[0013] Step S2.1, assuming the input is time domain data ,in , Indicates that the first dimension is , a real number tensor with the second dimension of 1; establish a long short-term memory network (LSTM) stacking module as a time domain feature extraction network module to extract time domain features and obtain the time domain features ,in , Represents the feature dimension at each time point.

[0014] Step S3, constructing a time-frequency domain image feature extraction network module, including the following steps:

[0015] Step S3.1, assuming the input is time-frequency domain image data ,in , Indicates the image height and width, Indicates the number of channels; establish a two-dimensional convolutional network based on channel and spatial attention mechanism. The two-dimensional convolutional network uses the convolutional block attention module CBAM based on channel attention and spatial attention mechanism to preliminarily extract the channel and spatial features of the time-frequency domain image data and obtain the shallow time-frequency domain image features. ,in , Represents shallow time-frequency domain image features The number of channels;

[0016] Step S3.2, based on the two-dimensional convolutional network in step S3.1, add a Transformer-Encoder module to form a time-frequency domain image feature extraction network module, and use the Transformer-Encoder module to extract the shallow time-frequency domain image features obtained in step S3.1 Encode and extract the deep features of the time-frequency domain image to obtain the time-frequency domain image features ,in , represents the number of features, Representing time-frequency image domain features The characteristic dimension of .

[0017] Step S4, constructing an interactive fusion module based on a cross-domain attention mechanism, including the following steps:

[0018] Step S4.1, extracting the time-frequency domain image features obtained by the time-frequency domain image feature extraction network module in step S3.2 And the time domain features obtained by the time domain feature extraction network module in step S2.1 Perform multi-head linear mapping and define the total number of heads as ; The first The time-frequency domain image features of each head are recorded as query , No. The time domain features of each head are recorded as keys Sum ,in , , and ;in, Representation query The characteristic dimension of Display key The characteristic dimension of Representation value The characteristic dimension of

[0019] Step S4.2, first, calculate the image features in the time-frequency domain The attention weights of each head generated under the guidance , according to the calculation method of the cross-domain attention mechanism, Weight of head The calculation formula is Then, calculate the Time-frequency domain image features in individual heads New temporal features to guide generation , as the output. Finally, the outputs of all heads are concatenated together and linearly transformed to obtain the time-frequency domain image features New temporal features generated under the guidance of :

[0020] ;

[0021] in, Display key The characteristic dimension of Indicates splicing; represents the linear change matrix; Indicates Size; Indicates New time domain features generated in each head;

[0022] Step S4.3, extracting the time-frequency image domain features obtained by the time-frequency domain image feature extraction network module in step S3.2 And the time domain features obtained by the time domain feature extraction network module in step S2.1 Perform multi-head linear mapping again and define the total number of heads as ; The first The time-frequency domain image features of each head are recorded as keys ,value , No. The time domain features of each head are recorded as query ,in , , and ;in, Display key The characteristic dimension of Representation query The characteristic dimension of Representation value The characteristic dimension of

[0023] Step S4.4, first, calculate the time domain feature The attention weights of each head generated under the guidance , according to the calculation method of the cross-domain attention mechanism, The weight calculation formula for each head is: Then, calculate the Time domain characteristics of each head New time-frequency domain image features generated under the guidance of Finally, the outputs of all heads are concatenated and linearly transformed to obtain the time domain features. New time-frequency domain image features generated under the guidance of :

[0024] ;

[0025] in, Display key The characteristic dimension of represents the linear change matrix; j represents the first Size; Indicates New time-frequency domain image features generated in each head;

[0026] Step S4.5, the new time domain features obtained in steps S4.2 and S4.4 are and time-frequency domain image features Perform global average pooling, then map it to a unified dimension through a fully connected layer, and concatenate to obtain deep interaction features. .

[0027] Step S5, based on the time domain feature extraction network module, the time-frequency domain image feature extraction network module and the interactive fusion module based on the cross-domain attention mechanism, a multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism is constructed, including the following steps:

[0028] Step S5.1, convert the time domain data in step S2.1 and the time-frequency domain image data in step S3.1 The time domain features are extracted by the time domain feature extraction network module and the time-frequency domain image feature extraction network module. and time-frequency domain image features ;

[0029] Step S5.2: Obtain new temporal features through an interactive fusion module based on a cross-domain attention mechanism and time-frequency domain image features , the features of the two domains guide each other to merge. and Send it to the global average pooling layer and the fully connected layer for processing and concatenation to obtain deep interaction features ;

[0030] Step S5.3: deep interaction features After the full connection layer and sigmiod activation function processing, the instability prediction probability is obtained ;

[0031] Step S5.4, the predicted probability of instability The improved cross entropy between the labels in step S1.2 is used as the loss function, and the loss function is as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] in, represents the loss function; represents the total number of samples; represents the category weight coefficient, , the greater the proportion of category 1 samples to all samples, the smaller its value; represents the prediction probability of the prediction model constructed in step S5, is the regulating factor, is the predicted probability of instability; represents the label of the tth sample;

[0036] Step S5.5, training the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism, saving the trained multi-source data fusion compressor instability prediction model and verifying it on the validation set, adjusting the hyperparameters of the multi-source data fusion compressor instability prediction model according to the validation set indicators, using the F-score score to quantify the prediction performance of the multi-source data fusion compressor instability prediction model, and then saving the optimal multi-source data fusion compressor instability prediction model based on the comprehensive performance indicator results, and finally obtaining the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism;

[0037] The F-score is calculated as follows:

[0038] ;

[0039] in, is the precision rate and R is the recall rate.

[0040] Step S6, preprocessing the test set data according to step S1, and adjusting the data dimension according to the input requirements of the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism, and finally obtaining the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism by step S5 for real-time prediction; the specific steps are as follows:

[0041] Step S6.1, in chronological order, using the multi-source data fusion compressor instability prediction model obtained in step S5.3 to give a surge prediction probability for each sample;

[0042] Step S6.2, select a set of dynamic pressure data, which includes experimental data recorded by a sensor that is working abnormally, and use the multi-source data fusion compressor instability prediction model obtained in step S5.3 to predict the stall probability of this set of data to test the fault tolerance of the multi-source data fusion compressor instability prediction model.

[0043] Furthermore, in the real-time prediction process: first, the test set data is shaped, and the data dimensions are adjusted according to the input requirements of the two feature extraction branch network models; in chronological order, a multi-source data fusion prediction model based on a cross-domain attention mechanism is used to give the surge prediction probability of each sample; a multi-source data fusion prediction model based on a cross-domain attention mechanism is used to give the stall prediction probability of the data recorded by a sensor with an abnormal operation.

[0044] Beneficial effects of the present invention:

[0045] (1) Compared with traditional methods, the method provided by the present invention integrates multi-source data information across fields, focuses on cross-field feature interaction, improves information completeness and prediction accuracy, and has a certain degree of fault tolerance.

[0046] (2) At the same time, the multi-source data fusion compressor instability prediction model outputs the surge prediction probability. The threshold value can be set to 0.4. By judging whether the surge prediction probability output by the prediction model is greater than 0.4, it is determined whether to alarm. The engine operating state is adjusted according to the surge probability. Therefore, the method of the present invention is conducive to improving the performance of engine active control.

[0047] (3) The present invention is based on dynamic pressure data and has nothing to do with the engine model. Therefore, by training different data sets, the multi-source data fusion compressor instability prediction model can be easily migrated to different models of engines for use, which has a certain degree of universality.

[0048] In summary, the present invention comprehensively processes and analyzes data from multiple sources, captures the complementarity and correlation between different data sources, forms a more comprehensive, complete and consistent perception of the target observation object, reduces diagnostic errors caused by a single data source, and improves the accuracy and reliability of prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the compressor instability prediction method based on multi-source data fusion;

[0050] Figure 2 This is the data preprocessing flow chart;

[0051] Figure 3 It is a structural diagram of the time domain feature extraction network module and the time-frequency domain image feature extraction network module;

[0052] Figure 4 It is the interactive fusion module;

[0053] Figure 5 The model structure diagram of the compressor instability prediction method based on multi-source data fusion;

[0054] Figure 6 This is the prediction result diagram of the multi-source data fusion compressor instability prediction method on the test set; Figure 6 (a) is the real-time change diagram of the dynamic pressure at the tip of the first-stage inlet stator. Figure 6 (b) is the real-time change diagram of surge prediction probability given by the multi-source data fusion compressor instability prediction method. Figure 6 (c) in the figure is the warning signal given by the multi-source data fusion compressor instability prediction model. DETAILED DESCRIPTION

[0055] The present invention is further described below in conjunction with the accompanying drawings. The present invention is based on a compressor gasping experiment of a certain type of aircraft engine. The process of the multi-source data fusion axial flow compressor instability prediction method provided by the present invention is as follows: Figure 1 As shown, the following steps are included:

[0056] Step S1, preprocessing the dynamic pressure sensor data in the compressor gasping experiment of a certain type of aircraft engine, dividing it into two types of data, including time domain data and time-frequency domain image data, the multi-source data in the present invention refers to the time domain data and the time-frequency domain image data, including the following steps:

[0057] Step S1.1, obtaining dynamic pressure sensor data in the compressor gasping experiment, including data in a stable state, an initial disturbance state, and an unstable state, and normalizing, filtering, and downsampling the dynamic pressure sensor data to obtain time domain data; specifically:

[0058] Figure 2 The data preprocessing flow chart shows that 10 measurement points were set in the experiment to measure the dynamic pressure value for t seconds from normal to surge. The sensor measurement frequency was 6kHz, and a total of 16 sets of data were recorded; the 10 measurement points were located at: the tip of the inlet guide vane stator, the tip of the zero-stage stator, the tip of the first-stage stator (three in the circumferential direction), the tip of the second-stage stator, the tip of the third-stage stator, the tip of the fourth-stage stator, the tip of the fifth-stage stator, and the outlet wall. The data preprocessing steps are as follows:

[0059] Step S1.1.1, normalizing the dynamic pressure sensor data in the compressor gasping experiment, and mapping the data distribution to the [0,1] interval through linear change;

[0060] Step S1.1.2, using an adaptive filter to filter the normalized dynamic pressure sensor data;

[0061] Step S1.1.3, in order to reduce the data dimension, the downsampling rate is determined according to the surge frequency distribution interval obtained by spectrum analysis and the dynamic pressure sensor data after adaptive filtering is downsampled to obtain time domain data;

[0062] Step S1.2, using the sliding segmentation method, the time domain data obtained in step S1.1 is segmented with a time step size of 128, and each segmented sample is labeled with 1 or 0 according to whether there is surge or not, wherein the label with surge is 1, and the label without surge is 0;

[0063] Step S1.3, in order to construct a multi-source data set and meet the requirements of time alignment, a wavelet analysis method is used to generate time-frequency domain image data for each sample in the time domain data set, and a label of 1 or 0 is marked according to whether each sample surges or not, where the label with surge is 1 and the label without surge is 0. The time domain data samples and the time-frequency domain image data samples in all time windows are paired to form a multi-source data set;

[0064] Step S1.4, the multi-source dataset consists of time domain data and time-frequency domain image data in pairs. The multi-source dataset is divided into a training dataset and a test dataset at a ratio of 6033:1500, and then the training dataset is divided into a training set and a validation set at a ratio of 4826:1207.

[0065] Step S2, constructing a time domain feature extraction network module, includes the following steps:

[0066] Step S2.1, assuming the input is time domain data ,in , Indicates that the first dimension is , a real tensor with the second dimension of 1; establish a long short-term memory network (LSTM) stacking module as a time domain feature extraction network module to extract time domain features and obtain the time domain features ,in , Represents the feature dimension at each time point.

[0067] Step S3, constructing a time-frequency domain image feature extraction network module, including the following steps:

[0068] Step S3.1, assuming the input is time-frequency domain image data ,in , Indicates the image height and width, Indicates the number of channels; establish a two-dimensional convolutional network based on channel and spatial attention mechanism. The two-dimensional convolutional network uses the convolutional block attention module CBAM based on channel attention and spatial attention mechanism to preliminarily extract the channel and spatial features of the time-frequency domain image data and obtain the shallow time-frequency domain image features. ,in , Represents shallow time-frequency domain image features The number of channels;

[0069] Step S3.2, based on the two-dimensional convolutional network in step S3.1, add a Transformer-Encoder module to form a time-frequency domain image feature extraction network module, and use the Transformer-Encoder module to extract the shallow time-frequency domain image features obtained in step S3.1 Encode and extract the deep features of the time-frequency domain image to obtain the time-frequency domain image features ,in , represents the number of features, Representing time-frequency image domain features In this embodiment, Figure 3This is the structure diagram of the time domain feature extraction network module and the time-frequency domain image feature extraction network module. The time domain feature extraction network module and the time domain image feature extraction network module are built for the two modalities of time domain data and time-frequency domain image data respectively.

[0070] Step S4: construct an interactive fusion module based on the cross-domain attention mechanism. Figure 4 It is an interactive fusion module based on the cross-domain attention mechanism. This step mainly performs cross-domain feature interactive fusion, as follows:

[0071] Time-frequency domain image features Guide the generation of new time-domain features:

[0072] Step S4.1, extracting the time-frequency domain image features obtained by the time-frequency domain image feature extraction network module in step S3.2 And the time domain features obtained by the time domain feature extraction network module in step S2.1 Perform multi-head linear mapping and define the total number of heads as ; Record the multi-head time-frequency domain image features as query , record the multi-head time domain features as keys Sum ; The first The time-frequency domain image features of each head are recorded as query , No. The time domain features of each head are recorded as keys Sum ,in , , and ;in, Representation query The characteristic dimension of Display key The characteristic dimension of Representation value The characteristic dimension of

[0073] Step S4.2, first, calculate the image features in the time-frequency domain The attention weights of each head generated under the guidance , according to the calculation method of the cross-domain attention mechanism, Weight of head The calculation formula is Then, calculate the Time-frequency domain image features in individual heads New temporal features to guide generation , as the output. Finally, the outputs of all heads are concatenated together and linearly transformed to obtain the time-frequency domain image features New temporal features generated under the guidance of :

[0074] ;

[0075] in, Display key The characteristic dimension of Indicates splicing; represents the linear change matrix; Indicates Size; Indicates New time domain features generated in each head; time domain features Guide the generation of new time-frequency domain image features;

[0076] Step S4.3, extracting the time-frequency image domain features obtained by the time-frequency domain image feature extraction network module in step S3.2 And the time domain features obtained by the time domain feature extraction network module in step S2.1 Perform multi-head linear mapping again and define the total number of heads as ; The multi-head time-frequency domain image features after multi-head linear mapping are recorded as the key Sum , and record the multi-head time domain features as query ; The first The time-frequency domain image features of each head are recorded as keys ,value , No. The time domain features of each head are recorded as query ,in , , and ;in, Display key The characteristic dimension of Representation query The characteristic dimension of Representation value The characteristic dimension of

[0077] Step S4.4, first, calculate the time domain feature The attention weights of each head generated under the guidance , according to the calculation method of the cross-domain attention mechanism, The weight calculation formula for each head is: Then, calculate the Time domain characteristics of each head New time-frequency domain image features generated under the guidance of Finally, the outputs of all heads are concatenated and linearly transformed to obtain the time domain features. New time-frequency domain image features generated under the guidance of :

[0078] ;

[0079] in, Display key The characteristic dimension of represents the linear change matrix; j represents the first Size; Indicates New time-frequency domain image features generated in each head;

[0080] Step S4.5, the new time domain features obtained in steps S4.2 and S4.4 are and time-frequency domain image features Perform global average pooling, then map it to a unified dimension through a fully connected layer, and concatenate to obtain deep interaction features. .

[0081] Step S5, constructing a multi-source data fusion compressor instability prediction model based on a cross-domain attention mechanism, Figure 5 This is the model structure diagram of the compressor instability prediction method based on multi-source data fusion and cross-domain attention mechanism. The construction steps are as follows:

[0082] Step S5.1: The time domain data after data preprocessing and time-frequency domain image data The time domain features are extracted by the time domain feature extraction network module and the time-frequency domain image feature extraction network module. and time-frequency domain image features ;

[0083] Step S5.2: Obtain new temporal features through an interactive fusion module based on a cross-domain attention mechanism and new time-frequency domain image features , the features of the two domains guide each other to merge. Then, the obtained and Send it to the global average pooling layer, the fully connected layer for mapping and concatenation to obtain deep interaction features ;

[0084] Step S5.3: deep interaction features After the full connection layer and sigmiod activation function processing, the instability prediction probability is obtained ;

[0085] Step S5.4, using the instability prediction probability The improved cross entropy between the labels is used as the loss function and is calculated as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] in, represents the loss function; represents the total number of samples; represents the category weight coefficient, , the greater the proportion of category 1 samples to all samples, the smaller its value; express Figure 5 The predicted probability of the prediction model shown is is the regulating factor, is the predicted probability of instability; represents the label of the tth sample;

[0090] Step S5.5, training the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism, saving the trained multi-source data fusion compressor instability prediction model and testing it on the validation set, adjusting the hyperparameters of the multi-source data fusion compressor instability prediction model according to the validation set indicators, using the F-score score to quantify the prediction performance of the multi-source data fusion compressor instability prediction model, and then saving the optimal multi-source data fusion compressor instability prediction model based on the comprehensive performance indicator results, and finally obtaining the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism;

[0091] The F-score is calculated as follows:

[0092] ;

[0093] in, is the precision rate and R is the recall rate.

[0094] Step S6, preprocessing the test set data according to step S1, and adjusting the data dimension according to the input requirements of the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism; the specific steps are as follows:

[0095] Step S6.1, in chronological order, using the multi-source data fusion compressor instability prediction model obtained in step S5.3 to give a surge prediction probability for each sample;

[0096] Step S6.2, select a set of dynamic pressure data, which includes experimental data recorded by a sensor that is working abnormally, and use the multi-source data fusion compressor instability prediction model obtained in step S5.3 to predict the stall probability of this set of data to test the fault tolerance of the multi-source data fusion compressor instability prediction model.

[0097] Figure 6 This is the prediction result diagram of the multi-source data fusion compressor instability prediction method on the test set, where Figure 6 (a) is the real-time change diagram of the dynamic pressure at the tip of the first-stage inlet stator. Figure 6 (b) is the real-time change diagram of surge prediction probability given by the multi-source data fusion compressor instability prediction method. Figure 6 (c) in the figure is the warning signal given by the multi-source data fusion compressor instability prediction model. Figure 6 As can be seen from Figure (a) in the figure, the dynamic pressure data shows an upward convex change at about 8s, indicating that the compressor has an unstable precursor feature. Figure 6 The predicted probability in (b) changes synchronously, and the predicted probability exceeds the threshold value 0.4 and is close to 1, which means that the surge prediction probability given by the present invention can successfully identify the occurrence of unstable precursors. Figure 6 In (c), a warning signal appears at the same time, indicating that the method proposed in the present invention can successfully achieve early warning.

[0098] The above-described embodiments merely express the implementation methods of the present invention, but they should not be understood as limiting the scope of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A multi-source data fusion axial flow compressor instability prediction method, characterized in that: The multi-source data fusion axial flow compressor instability prediction method comprises the following steps: Step S1, preprocessing the dynamic pressure sensor data in the compressor gasping experiment of a certain type of aircraft engine, dividing the data into two types of data, including time domain data and time-frequency domain image data, obtaining a multi-source data set, and dividing the data set into a test data set, a training set, and a validation set; Step S2, constructing a time domain feature extraction network module; Step S3, constructing a time-frequency domain image feature extraction network module; Step S4, constructing an interactive fusion module based on a cross-domain attention mechanism; Step S5, constructing a multi-source data fusion compressor instability prediction model based on a cross-domain attention mechanism based on a time domain feature extraction network module, a time-frequency domain image feature extraction network module and an interactive fusion module based on a cross-domain attention mechanism; training the model and saving the optimal model; Step S6, preprocess the test set data according to step S1, and perform real-time prediction using the optimal model obtained in step S5.

2. The multi-source data fusion axial flow compressor instability prediction method according to claim 1 is characterized in that: The step S1 comprises the following steps: Step S1.1, obtaining dynamic pressure sensor data in the compressor gasping experiment, including data in a stable state, an initial disturbance state, and an unstable state, and normalizing, filtering, and downsampling the dynamic pressure sensor data to obtain time domain data; Step S1.2, segmenting the time domain data obtained in step S1.1 according to the sliding segmentation method, and labeling each segmented sample with 1 or 0 according to whether there is surge or not, wherein the label with surge is 1, and the label without surge is 0; Step S1.3, using the wavelet analysis method to generate the time-frequency domain image data of the corresponding time step, and labeling each sample after segmentation with 1 or 0 according to whether there is surge or not, wherein the label with surge is 1, and the label without surge is 0; Step S1.4, the multi-source data set consists of time domain data and time-frequency domain image data in pairs. The data set is divided into a training data set and a test data set in proportion, and then the training data set is divided into a training set and a validation set.

3. The method for predicting axial compressor instability by multi-source data fusion according to claim 2 is characterized by: The step S2 comprises the following steps: Step S2.1, assuming the input is time domain data ,in , Indicates that the first dimension is , the second dimension is a real tensor of 1; establish a long short-term memory network stacking module as a time domain feature extraction network module to extract time domain features and obtain time domain features ,in , Represents the feature dimension at each time point; The step S3 comprises the following steps: Step S3.1, assuming the input is time-frequency domain image data ,in , Indicates the image height and width, Indicates the number of channels; establishes a two-dimensional convolutional network based on channel and spatial attention mechanism; the two-dimensional convolutional network uses the convolutional block attention module CBAM based on channel attention and spatial attention mechanism to preliminarily extract the channel and spatial features of the time-frequency domain image data to obtain shallow time-frequency domain image features ,in , Represents shallow time-frequency domain image features The number of channels; Step S3.2, based on the two-dimensional convolutional network in step S3.1, add a Transformer-Encoder module to form a time-frequency domain image feature extraction network module, and use the Transformer-Encoder module to extract the shallow time-frequency domain image features obtained in step S3.1 Encode and extract the deep features of the time-frequency domain image to obtain the time-frequency domain image features ,in , represents the number of features, Representing time-frequency image domain features The characteristic dimension of .

4. The method for predicting axial compressor instability by multi-source data fusion according to claim 3 is characterized in that: The step S4 comprises the following steps: Step S4.1, extracting the time-frequency domain image features obtained by the time-frequency domain image feature extraction network module in step S3.2 And the time domain features obtained by the time domain feature extraction network module in step S2.1 Perform multi-head linear mapping and define the total number of heads as ; The first The time-frequency domain image features of each head are recorded as query , No. The time domain features of each head are recorded as keys Sum ,in , , and ;in, Representation query The characteristic dimension of Display key The characteristic dimension of Representation value The characteristic dimension of Step S4.2, first, calculate the image features in the time-frequency domain The attention weights of each head generated under the guidance , according to the calculation method of the cross-domain attention mechanism, Weight of head The calculation formula is Then, calculate the Time-frequency domain image features in individual heads New temporal features to guide generation , as the output; finally, the outputs of all heads are spliced ​​together and linearly transformed to obtain the time-frequency domain image features New temporal features generated under the guidance of : ； in, Display key The characteristic dimension of Indicates splicing; represents the linear change matrix; Indicates Size; Indicates New time domain features generated in each head; Step S4.3, extracting the time-frequency image domain features obtained by the time-frequency domain image feature extraction network module in step S3.2 And the time domain features obtained by the time domain feature extraction network module in step S2.1 Perform multi-head linear mapping again and define the total number of heads as ; The first The time-frequency domain image features of each head are recorded as keys ,value , No. The time domain features of each head are recorded as query ,in , , and ;in, Display key The characteristic dimension of Representation query The characteristic dimension of Representation value The characteristic dimension of Step S4.4, first, calculate the time domain feature The attention weights of each head generated under the guidance , according to the calculation method of the cross-domain attention mechanism, The weight calculation formula for each head is: ; Then, calculate the time domain features in the jth head New time-frequency domain image features generated under the guidance of , get the output; finally, concatenate the outputs of all heads together and perform linear transformation to get the time domain features New time-frequency domain image features generated under the guidance of : ； in, Display key The characteristic dimension of represents the linear change matrix; j represents the first Size; Indicates New time-frequency domain image features generated in each head; Step S4.5, the new time domain features obtained in steps S4.2 and S4.4 are and time-frequency domain image features Perform global average pooling, then map it to a unified dimension through a fully connected layer, and concatenate to obtain deep interaction features. .

5. The method for predicting axial compressor instability by multi-source data fusion according to claim 4 is characterized in that: The step S5 comprises the following steps: Step S5.1, convert the time domain data in step S2.1 and the time-frequency domain image data in step S3.1 The time domain features are extracted by the time domain feature extraction network module and the time-frequency domain image feature extraction network module. and time-frequency domain image features ; Step S5.2: Obtain new temporal features through an interactive fusion module based on a cross-domain attention mechanism and time-frequency domain image features , the features of the two domains guide each other to merge; the obtained and Send it to the global average pooling layer and the fully connected layer for processing and concatenation to obtain deep interaction features ; Step S5.3: deep interaction features After the full connection layer and sigmiod activation function processing, the instability prediction probability is obtained ; Step S5.4, the predicted probability of instability The improved cross entropy between the labels in step S1.2 is used as the loss function; Step S5.5, train the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism, save the trained multi-source data fusion compressor instability prediction model and verify it on the validation set, adjust the hyperparameters of the multi-source data fusion compressor instability prediction model according to the validation set indicators, use the F-score score to quantify the prediction performance of the multi-source data fusion compressor instability prediction model, save the optimal multi-source data fusion compressor instability prediction model based on the comprehensive performance indicator results, and finally obtain the multi-source data fusion compressor instability prediction model based on the cross-domain attention mechanism.

6. The multi-source data fusion axial flow compressor instability prediction method according to claim 5 is characterized in that: In step S5.4, the loss function is as follows: ； ； ； in, represents the loss function; represents the total number of samples; represents the category weight coefficient, , the greater the proportion of category 1 samples to all samples, the smaller its value; represents the prediction probability of the prediction model constructed in step S5, is the regulating factor, is the predicted probability of instability; Represents the label of the tth sample.

7. The method for predicting axial compressor instability by multi-source data fusion according to claim 5, characterized in that: In step S5.4, the F-score is calculated as follows: ； in, is the precision rate and R is the recall rate.

8. The method for predicting axial compressor instability by multi-source data fusion according to claim 5, characterized in that: The step S6 comprises the following steps: Step S6.1, in chronological order, using the multi-source data fusion compressor instability prediction model obtained in step S5.3 to give a surge prediction probability for each sample; Step S6.2, select a set of dynamic pressure data, which includes experimental data recorded by a sensor that is working abnormally, and use the multi-source data fusion compressor instability prediction model obtained in step S5.3 to predict the stall probability of this set of data to test the fault tolerance of the multi-source data fusion compressor instability prediction model.

Images