Aviation compressor stall surge online identification method based on deep learning

Abstract

The invention discloses a deep learning-based online identification method for stall and surge of an aviation compressor, and belongs to the field of aviation compressor stability research. A three-layer architecture is adopted, wherein a data layer collects dynamic pressure signals of a gas compressor in real time, and 512-dimensional feature vectors are obtained through FFT preprocessing; the logic layer constructs and trains a CNN model, an LSTM model and a CNN-GRU model, and an optimal model is screened out; and the presentation layer realizes interaction and visual presentation of upper and lower computers through LabVIEW. According to the method, stall surge can be pre-warned more than 50 ms in advance, manual dependence is reduced, embedded deployment is adapted, and the running safety and economical efficiency of the aero-engine can be improved.

Description

Technical Field

[0001] This invention belongs to the field of fault detection technology for aircraft compressors, specifically, it relates to an online identification method for stall and surge in aircraft compressors based on deep learning. Background Technology

[0002] The aerodynamic stability issues of high-performance aero compressors primarily originate from the compressor itself. The compressor's operating condition and stability are crucial to the overall engine efficiency and safety, and predicting its instability has always been a research hotspot and challenge in the aero-engine field. As a core component of an aero-engine, the compressor improves system efficiency by adjusting the airflow pressure ratio. However, when the pressure ratio exceeds the surge threshold, the compressor is prone to rotating stall, leading to airflow separation, sudden pressure drops, and severe vibrations. If such faults are not detected in time, they can cause turbine blade damage, abnormally high combustion chamber temperatures, and even irreversible engine damage. Therefore, it is necessary to monitor pressure signals in real time using sensors and trigger anti-surge control measures (such as guide vane adjustment and intermediate stage bleed-out) in advance to prevent the stall evolution process.

[0003] Compared to traditional shallow neural networks, deep learning features a significantly more complex structure with a substantial increase in the number of hidden layers, thus possessing stronger nonlinear modeling capabilities and the ability to extract higher-order abstract features from data layer by layer. The increased depth leads to an exponential growth in the number of network parameters with each layer, necessitating the optimization of weight distribution using massive amounts of training samples to enhance model generalization performance. Deep learning can autonomously uncover complex mapping relationships between inputs and outputs, dynamically adjusting network parameters through gradient backpropagation to construct accurate representations of system behavior. The completeness and robustness of these representations directly determine the accuracy of fault diagnosis.

[0004] Convolutional Neural Networks (CNNs) are primarily used for image feature extraction. They consist of convolutional layers, pooling layers, and fully connected layers. Convolutional layers are the foundation of feature extraction, with the convolutional kernel reflecting local image features. A pooling layer is typically added between adjacent convolutional layers in a CNN. Pooling layers effectively reduce the size of the parameter matrix, thereby reducing the number of parameters in the final connected layers. Therefore, adding pooling layers can speed up computation and prevent overfitting. The fully connected layers are located at the end of the CNN network. After feature extraction by the convolutional layers and dimensionality reduction by the pooling layers, the feature maps are converted into one-dimensional vectors and fed into the fully connected layers for classification or regression operations.

[0005] Recurrent Neural Networks (RNNs) are artificial neural networks specifically designed to process sequential data. Unlike traditional feedforward neural networks, RNNs possess "memory," allowing them to utilize past information to influence current outputs. They are connected not only between layers but also between nodes within the same layer, forming a loop. This loop allows information to propagate from one time step to the next. Each training sample in a RNN is a time series, with correlations between the input values ​​of the same training sample at different times. The sequence length for each sample may vary. During training, forward propagation is performed on the input values ​​at each time step in the sequence, followed by backpropagation to calculate the gradient values ​​of the parameters and update them. This structure enables RNNs to handle sequential data of arbitrary length. Deep learning algorithms based on CNNs and RNNs are suitable for identifying and predicting compressor stall pressure signals.

[0006] Therefore, this invention provides an online identification method for stall surge of aero-engine compressor based on deep learning, which greatly reduces the occurrence of unstable operation of the compressor and achieves high-precision stall fault diagnosis. Summary of the Invention

[0007] In view of this, the technical problem to be solved by the present invention is to provide an online identification method for stall surge of aero-engine compressor based on deep learning. By constructing a three-layer architecture of "data layer - logic layer - representation layer", the method can realize real-time and high-precision early warning of compressor instability conditions, reduce the risk of unstable operation, and improve the safety of engine operation.

[0008] To address the aforementioned technical problems, this invention discloses an online identification method for stall and surge in aero-engine compressors based on deep learning, comprising the following steps: The specific implementation steps of this method are as follows: The first step is data acquisition and preprocessing. For data acquisition, real-time dynamic pressure signals from the aircraft compressor are collected. The acquisition location is a cross-section of the leading edge of the compressor rotor, where pressure sensors are evenly arranged along the circumference. These sensors are spaced at 45° intervals along the circumference of the compressor rotor's leading edge cross-section, covering the pressure values ​​under all operating conditions within the compressor's surge boundary. The sensors are fixed to a steel frame platform with a 33mm axial clearance from the compressor test piece. During acquisition, data from two symmetrically positioned measurement points out of six are selected, and the sensor sampling frequency is 25600Hz. In the preprocessing stage, the acquired raw pressure signal is processed by using a recursive method to decompose the signal sequence, transforming the even-numbered and odd-numbered parts separately, and performing butterfly merging through a rotation factor to obtain the spectrum of the time-domain signal. A Fast Fourier Transform is then performed on the real-number input signal to extract the first half of the spectrum. The normalized feature of the spectrum, expressed as log-magnitude, is calculated with a normalization standard of zero mean and unit standard deviation, returning a 512-dimensional normalized feature vector. Finally, the preprocessed signal is segmented into time windows, and the segmented data is used for subsequent model training and online identification.

[0009] The second step is model training and selection at the logic layer. This layer requires building and training CNN, LSTM, and CNN-GRU models. The optimal model is selected by comparing the warning performance of these models under different speed conditions. Warning performance includes response time and accuracy. These models also have functions for structure initialization, parameter optimization, forward propagation, backpropagation update, and model saving and import. Specifically, the input layer of the CNN model receives a 512-dimensional feature vector; the convolutional layer uses eight 5×1 kernels with ReLU activation function, and the initial weights are uniformly distributed in the range [-1,1]; the pooling layer uses max pooling with a window size of 2×1, retaining the maximum value position information for backpropagation; the fully connected layer calculates the loss using cross-entropy combined with the Softmax function. During backpropagation, the fully connected layer directly calculates the weight gradient, while the pooling layer only propagates the gradient to the maximum value position. The LSTM model consists of one LSTM unit and one fully connected output layer. The LSTM unit is composed of a forget gate, an input gate, an output gate, and candidate memories. The input dimension is 512, the hidden layer size is 16, the learning rate is 0.0001, the binary classification output dimension is 2, and the batch size is 32. The weights and biases are initialized with a uniform random distribution in the range of [-0.1, 0.1]. During forward propagation, the gate value, memory state, and hidden state of each time step are saved to the buffer for backpropagation. Finally, the output is calculated through the hidden state and the probability distribution is obtained by the Softmax function. The CNN-GRU model consists of a CNN feature extraction module and a GRU sequence modeling module connected in series. The CNN module takes 16-channel features as input, extracts features through convolutional layers, and outputs an 8-channel feature vector. The convolutional kernels in the convolutional layers are 5×1 in size and number 8. The GRU module receives the temporal feature vector output by the CNN, has 16 hidden layer neurons, captures long-term dependencies in the sequence through update and reset gates, and finally outputs the classification result through a fully connected layer.

[0010] The model training process is as follows: First, the dataset is divided. The preprocessed 512-dimensional feature vectors are divided into training and test sets in a 7:3 ratio. The training set is then divided into sub-training and validation sets in a 4:1 ratio. Next, model training is performed, constructing three models. Mini-batch gradient descent is used to optimize parameters. The parameters for this method are a momentum coefficient of 0.9, a weight decay coefficient of 1e-5, and an initial learning rate of 0.001. The learning rate is automatically decayed when the validation set loss stagnates, with a decay factor of 0.5. The loss function for CNN and CNN-GRU is cross-entropy + Softmax, while the loss function for LSTM is cross-entropy. During training, the sub-training set loss is calculated in real-time, and the weights are updated in reverse. After each training round, accuracy, precision, and response time are evaluated on the validation set. Training stops when there is no improvement in validation set performance after five consecutive rounds. The optimal model parameters on the validation set are saved as a binary file. Finally, the models were compared. The warning response time and accuracy of the three models were tested at three speed conditions: 1000rpm, 1200rpm, and 1400rpm. The model with the best overall performance was selected as the main online identification model.

[0011] Finally, there is the online identification and interaction of the presentation layer. The trained optimal model is integrated into the LabVIEW platform via the extern "C" interface to build a visual interface that includes sensor signal display, early warning result display, and parameter adjustment functions. Real-time data transmission and inference are achieved through network communication between the upper and lower computers to perform online identification and early warning of compressor stall and surge conditions. Specifically, this includes building an upper and lower computer interaction system. The upper computer deploys a LabVIEW development environment and includes a visual interface, which includes three canvas areas displaying sensor 1 signals, sensor 2 signals, and early warning results, as well as four interactive modules: sensor coefficient calibration, threshold selection, early warning indicator, and a stop operation button. The lower computer uses an NI cRIO-9033 controller with an integrated FPGA module for high-speed signal acquisition, and a real-time processor that loads the cross-compiled model dynamic link library and parameter files. The real-time identification process involves the lower-level machine acquiring pressure signals at 25600Hz via an FPGA module. After preprocessing to generate a 512-dimensional feature vector, the signal is input into a pre-trained CNN-GRU model. The model outputs a stall and surge probability value. When the probability value exceeds a preset threshold of 0.85, the lower-level machine sends a warning signal to the upper-level machine via TCP / IP protocol. The upper-level machine interface triggers a warning indicator and records the warning time. Regarding model deployment and adaptation, the deep learning model exposes exported functions through the `extern "C"` interface. After cross-compiling with a Xilinx compiler, it is adapted to the embedded environment, supporting a real-time inference latency of ≤50ms on the NI cRIO-9033 controller.

[0012] Compared with the prior art, the present invention can achieve the following technical effects: (1) The stall fault diagnosis algorithm used in this invention can be trained by using historical operating data of aero-engines. The optimal model is selected by comparing the prediction results of different algorithm models, which greatly reduces the need for human experience and realizes a high-precision stall fault diagnosis and classification algorithm.

[0013] (2) This invention uses LabVIEW to implement the visualization and user interaction interface and supports online access to sensor devices for real-time analysis of pressure data and instability warning. In addition, the model is exposed to external programs through the extern "C" interface and can be encapsulated through cross-compilation toolchain, making it suitable for embedded deployment.

[0014] Of course, any product implementing this invention does not necessarily need to achieve all of the technical effects described above at the same time. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic diagram of the compressor instability prediction system architecture; Figure 2 This is a schematic diagram of the visual display interface; Figure 3 A simplified diagram of the compressor test bench system; Figure 4 This is a schematic diagram of the sensor arrangement. Figure 5 This is a schematic diagram showing the positions of various sections of the compressor; Figure 6 FFT images of dynamic pressure signals under speed conditions of 1000 rpm, 1200 rpm, and 1400 rpm; Figure 7 This is a schematic diagram of the prediction display interface for each algorithm model at 1000rpm; Figure 8 This is a schematic diagram of the prediction display interface for each algorithm model at 1200rpm; Figure 9 This is a schematic diagram of the prediction display interface for each algorithm model at 1400rpm; Figure 10 This is a schematic diagram of the results of the analysis of variance under different rotational speeds. Detailed Implementation

[0016] The following will describe in detail the implementation of the present invention with reference to the accompanying drawings and embodiments, so that the process of how the present invention uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0017] This invention discloses an online stall surge identification method for aero-engine compressors based on deep learning, belonging to the field of aero-engine compressor stability research. It achieves real-time early warning through a three-layer architecture of "data layer - logic layer - presentation layer," identifying stall surge more than 50ms in advance, reducing reliance on manual intervention, adapting to embedded deployment, and improving the operational safety and economy of aero-engines. Its core steps and technical features are as follows.

[0018] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the compressor instability prediction system architecture. It comprises a three-layer architecture: the first layer is the data layer, used to acquire dynamic pressure signals from the compressor and perform data preprocessing, time window segmentation, and FFT frequency domain feature extraction; the second layer is the logic layer, which uses artificial intelligence algorithms to provide early warnings of instability signals. Three different algorithm models are trained and compared to achieve the best prediction results, including CNN, LSTM, and CNN-GRU models; the third layer is the presentation layer, responsible for the user interface (GUI). It applies trained deep learning algorithm models in the LabVIEW graphical language to display early warning results and related information to the user, while simultaneously reading user actions and feeding back the corresponding data and logic layers.

[0019] Specifically, the following steps are included: S1: Data layer acquisition and preprocessing: Real-time acquisition of dynamic pressure signals from the aircraft compressor, FFT frequency domain feature extraction, time window segmentation and normalization preprocessing of the signals, and output of feature vectors; S2: Logical Layer Model Training and Selection: Construct and train CNN, LSTM, and CNN-GRU models. Select the optimal model by comparing the warning performance of the models under different speed conditions. Warning performance includes response time and accuracy. The model has functions for structure initialization, parameter optimization, forward propagation, backward update, and model saving and import. S3: Online Identification and Interaction of Presentation Layer: The trained optimal model is integrated into the LabVIEW platform through the extern "C" interface to build a visual interface that includes sensor signal display, early warning result display and parameter adjustment functions. Real-time data transmission and inference are realized through upper and lower computer network communication to identify and warn of compressor stall surge conditions online.

[0020] In step S3, the visualization section is as follows: Figure 2As shown, the deep learning model of this invention includes an exported model module, which is used to build, train, save, and import the deep learning model. The exported functions are exposed to external programs through the `extern "C"` interface and can be encapsulated using a cross-compilation toolchain for embedded deployment. The test system consists of two parts: a host computer and a slave computer. They work collaboratively via network communication to complete key tasks such as data acquisition, algorithm inference, and system control. First, LabVIEW, LabVIEW Real-Time, LabVIEW FPGA, NI-RIO drivers, and the Xilinx compiler integrated into the LabVIEW FPGA are deployed on the target PC, all requiring the same version. Simultaneously, the firmware in the NI MAX software is updated to the same version as the host computer software. On the Ubuntu 20.04 system, the artificial intelligence algorithm file implemented in C++ is compiled into a dynamic link library file using a cross-compilation toolchain. This file has cross-platform deployment capabilities and can be directly called from the target slave computer. The slave computer is connected in NI MAX, and the .so and .bin files are transferred to the slave computer directory. The visualization interface includes three canvas areas (dynamic pressure sensor 1 signal plotting, sensor 2 signal plotting, and warning result plotting). The right side contains four sections, from top to bottom: sensor 1 coefficient calibration, sensor 2 coefficient calibration, threshold selection, warning indicator display, and a stop operation button. Users can flexibly adjust parameters such as sensor coefficients.

[0021] Furthermore, step S1 specifically includes: S11: Real-time acquisition of compressor dynamic pressure signals. The acquisition location is on the cross-section of the compressor rotor leading edge. Pressure sensors are evenly arranged along the circumference. Data from two symmetrical measurement points out of six measurement points are selected. The sensor sampling frequency is 25600Hz. S12: Preprocess the acquired raw pressure signal, decompose the signal sequence using a recursive method, transform the even-numbered and odd-numbered parts separately, and complete the butterfly merging through a rotation factor to obtain the spectrum of the time-domain signal; perform a fast Fourier transform on the real-number input signal, extract the first half of the spectrum, calculate the normalized feature of log-magnitude on the spectrum, with the normalization standard being zero mean and unit standard deviation, and return a 512-dimensional normalized feature vector; S13: Perform time window segmentation on the preprocessed signal. The segmented data is used for subsequent model training and online identification.

[0022] In step S11, the present invention uses pressure test data from a single-stage compressor test bench to test the feasibility of the instability prediction model. The test bench consists of multiple components, including an inlet bell, an intake pipe, a fairing, a compressor test piece, an exhaust volute, an exhaust pipe, a throttle valve, a drive motor, and a support trolley. The test bench system is as follows: Figure 3 As shown in the diagram. High-performance sensors are deployed at the inlet and outlet of each stage of the compressor to collect dynamic pressure signals in real time. These signals are then converted into digital signals that a computer can recognize and input into a pre-built model. The model outputs and displays the current compressor health indicators and provides early warning signals in the event of a stall. The sensor deployment method is as follows: Figure 4 As shown, pressure sensors are evenly arranged along the circumference on the cross-section of the leading edge of the compressor rotor. The positions of each cross-section of the compressor are as follows: Figure 5 As shown in the figure. The interval between each measurement point is 45°, the pressure sensor sampling frequency is 25600Hz, and experiments were conducted at compressor speeds of 1000rpm, 1200rpm, and 1400rpm. The rotational stall process was simulated by continuously adjusting the opening of the throttle valve. In the experiment, each component was fixed to a steel frame platform by a bracket, and the compressor was rotated by a power unit. After entering the compressor specimen body, the air was discharged into the atmosphere at the end of the specimen. The configuration and design parameters of the compressor are shown in Table 1:

[0023] The original data for the model of this invention uses two symmetrical points out of six measurement points.

[0024] In step S12, the pressure data is first preprocessed using FFT. Specifically, the signal sequence is decomposed using a recursive method, and the even / odd parts are transformed separately. Finally, a butterfly merge is completed using a rotation factor to obtain the spectrum of the time-domain signal. Then, a Fast Fourier Transform is performed on the real-valued input signal to extract the first half of the spectrum. The normalized log-magnitude feature (zero mean, unit standard deviation) of the spectrum is calculated, and the returned normalized feature vector is obtained. The preprocessed output data contains 512-dimensional features.

[0025] In step S2, the input layer of the CNN model receives a 512-dimensional feature vector; the convolutional layer uses 8 convolutional kernels of size 5×1, with ReLU as the activation function, and the weights are initialized to be uniformly distributed in the range [-1,1]; the pooling layer uses max pooling with a window size of 2×1, retaining the maximum value position information for backpropagation; the fully connected layer calculates the loss by combining cross-entropy with the Softmax function, and during backpropagation, the fully connected layer directly calculates the weight gradient, while the pooling layer only passes the gradient to the maximum value position.

[0026] The LSTM model consists of one LSTM unit and one fully connected output layer. The LSTM unit is composed of a forget gate, an input gate, an output gate, and candidate memories. The input dimension is 512, the hidden layer size is 16, the learning rate is 0.0001, the binary classification output dimension is 2, and the batch size is 32. The weights and biases are initialized with a uniform random distribution in the range of [-0.1, 0.1]. During forward propagation, the gate value, memory state, and hidden state of each time step are saved to the buffer for backpropagation. Finally, the output is calculated through the hidden state and the probability distribution is obtained by the Softmax function.

[0027] The CNN-GRU model consists of a CNN feature extraction module and a GRU sequence modeling module connected in series. The CNN module takes 16-channel features as input, extracts features through convolutional layers, and outputs an 8-channel feature vector. The convolutional kernels in the convolutional layers are 5×1 in size and number 8. The GRU module receives the temporal feature vector output by the CNN, has 16 hidden layer neurons, captures long-term dependencies in the sequence through update and reset gates, and finally outputs the classification result through a fully connected layer.

[0028] The model training process is as follows: First, the dataset is divided. The preprocessed 512-dimensional feature vectors are divided into training and test sets in a 7:3 ratio. The training set is then divided into sub-training and validation sets in a 4:1 ratio. Next, model training is performed, constructing three models. Mini-batch gradient descent is used to optimize parameters. The parameters for this method are a momentum coefficient of 0.9, a weight decay coefficient of 1e-5, and an initial learning rate of 0.001. The learning rate is automatically decayed when the validation set loss stagnates, with a decay factor of 0.5. The loss function for CNN and CNN-GRU is cross-entropy + Softmax, while the loss function for LSTM is cross-entropy. During training, the sub-training set loss is calculated in real-time, and the weights are updated in reverse. After each training round, accuracy, precision, and response time are evaluated on the validation set. Training stops when there is no improvement in validation set performance after five consecutive rounds. The optimal model parameters on the validation set are saved as a binary file. Finally, the models were compared. The warning response time and accuracy of the three models were tested at three speed conditions: 1000rpm, 1200rpm, and 1400rpm. The model with the best overall performance was selected as the main online identification model.

[0029] Specifically, the raw pressure signal, dynamic pressure signal, and Fourier transform spectrum of a single-stage axial compressor at various speeds are as follows: Figure 6 As shown. Among them. Figure 6 (a) is a schematic diagram of measuring point 1 at 1000 rpm. Figure 6 (b) is a schematic diagram of FFT at 1000 rpm. Figure 6(c) is a schematic diagram of measuring point 1 at 1200 rpm. Figure 6 (d) is a schematic diagram of FFT at 1200 rpm. Figure 6 (e) is a schematic diagram of measuring point 1 at 1400 rpm. Figure 6 (f) is a schematic diagram of FFT at 1400rpm.

[0030] First, the recognition performance of three artificial intelligence algorithms was tested at 1000 rpm. The data length of the input algorithms was fixed at 1024. All three models were trained using offline data, and the optimal model parameters under single-condition conditions have been saved to a binary file. The prediction display interface of each model is shown below. Figure 7 ,in Figure 7 (a) is a diagram of a CNN model. Figure 7 (b) is a diagram of the LSTM model. Figure 7 (c) is a diagram of the CNN+GRU model.

[0031] The CNN model achieved stable identification of the compressor's pre-stall state at this RPM, with an average system response time of 75ms. It did not exhibit false alarms during normal operation and effectively identified compressor instability. The LSTM model performed well in processing time-series data, with a slightly longer response time of 92ms, but it could provide effective warnings before instability occurred. The CNN-GRU model, combining the spatial feature extraction capabilities of CNNs with the time-series modeling advantages of GRUs, achieved the best overall performance with a response time of 68ms.

[0032] Secondly, the algorithm's recognition performance at 1200 rpm was examined, such as... Figure 8 ,in Figure 8 (a) is a diagram of a CNN model. Figure 8 (b) is a diagram of the LSTM model. Figure 8 (c) shows the CNN+GRU model. The CNN model maintains stable performance at this rotational speed, with a response time of 73ms, and the algorithm output matches the actual surge event well. The LSTM model shows slight fluctuations in handling data feature perturbations caused by this speed change. The CNN-GRU model continues to demonstrate excellent performance and has not made any misclassifications in multiple tests.

[0033] Finally, the algorithm's recognition performance at 1400 rpm was tested, as shown below. Figure 9 ,in Figure 9 (a) is a diagram of a CNN model. Figure 9 (b) is a diagram of the LSTM model. Figure 9(c) shows the CNN+GRU model. The CNN model's adaptability to compressor state changes slightly decreases at high speeds, with a response latency of 79ms, but its identification still has reference value. The LSTM model performs relatively stably on high-frequency data, with a response time of 90ms, demonstrating good robustness to data fluctuations. The CNN-GRU model maintains its leading response rate, with a response time controlled within 71ms, exhibiting the best overall performance.

[0034] The test results under three operating speed conditions show that all three artificial intelligence models can effectively identify and warn of compressor stall and surge, meeting both real-time and accuracy requirements. The CNN model is suitable for scenarios with stable features and high real-time requirements; the LSTM model is suitable for data sequences with strong temporal dynamics; and the CNN-GRU model outperforms other models in both accuracy and response speed, making it the recommended primary algorithm in the current system. All models met the algorithm performance indicators specified in the project during testing and are capable of being deployed and operated in actual compressor systems. This demonstrates that the predictive model mentioned in this invention, when applied to the compressor in question, can provide a warning at least 50ms after stall occurs, and the faster the operating speed, the better the stall warning effect. It exhibits good performance in terms of real-time performance, stability, and accuracy of stall warnings and can adapt to pressure data under different operating speed conditions.

[0035] During real-time prediction, the communication mechanism between the upper and lower level uses a TCP / IP network connection, supporting the NI-RIO communication protocol to ensure real-time performance and data integrity. Data acquisition and uploading employ a timed trigger mechanism with configurable sampling periods. During LabVIEW program deployment, dependency loading and synchronization are automatically completed through the NI Deployment Utility. The upper computer uses a Windows operating system and runs the LabVIEW development environment. The developed application code is deployed to the CompactRIO controller on the lower computer via the LabVIEW programming environment. Using NI MAX tools or the FTP / SFTP protocol, the algorithm library files (.so files) and model parameter files (.bin files) are transferred to the LabVIEW directory "home / lvuser / ". The upper computer provides a visual interface and human-computer interaction, enabling monitoring of the running status, logging, and result display. It also receives real-time sampled data uploaded from the lower computer for display and post-processing.

[0036] The lower-level controller uses an NI cRIO-9033 embedded real-time controller, equipped with a real-time processor and a reconfigurable FPGA (Field-Programmable Gate Array). The lower-level controller acquires real-time signals during compressor operation at high speed via the FPGA module. It loads and calls the transmitted algorithm .so dynamic link library on the real-time processor to execute the inference logic. It loads model weights using a .bin parameter file to complete specific intelligent recognition or prediction tasks. The algorithm inference results and raw sensor data are transmitted back to the upper-level controller in real time for interface display or further analysis.

[0037] As an analogy, the analysis of heterogeneous variance (HPV) algorithm is used to analyze and provide early warnings of pressure data under the same compressor operating conditions. HPPV is a traditional instability diagnosis algorithm that uses the standard deviation of signals from two adjacent measurement points as a basis to calculate an anomaly index, which is then compared with a set threshold to determine whether the system is unstable. Combined with tools such as filter settings and windowed analysis, HPPV generally has a good effect on early warning of compressor instability. Its prediction results are as follows: Figure 10 As shown, where Figure 10 (a) is a warning information diagram under the condition of 1000 rpm. Figure 10 (b) is a warning information diagram under the condition of 1200 rpm. Figure 10 (c) shows the warning information at 1400 rpm. Compared with traditional anisotropic analysis of variance algorithms, the deep learning model of this invention can issue warnings more than 50 ms before stall occurs at 1000 rpm, 1200 rpm, and 1400 rpm, while traditional algorithms can only identify stall at the time of occurrence. The CNN-GRU model exhibits the best overall performance, with more stable warning effects at higher speeds, verifying the significant advantages of this method in real-time performance, accuracy, and robustness.

[0038] In summary, this invention achieves high-precision online identification of stall surge in aero-engine compressors through the integration of a three-layer architecture design and deep learning technology. Its technical details and implementation process form a complete closed loop and have clear engineering application value.

[0039] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A deep learning-based online identification method for stall and surge in aero-engine compressors, characterized in that, Includes the following steps: S1: Data layer acquisition and preprocessing: Real-time acquisition of dynamic pressure signals from the aircraft compressor, FFT frequency domain feature extraction, time window segmentation and normalization preprocessing of the signals, and output of feature vectors; S2: Logical Layer Model Training and Selection: Construct and train CNN, LSTM, and CNN-GRU models. Select the optimal model by comparing the early warning performance of the models under different speed conditions. The early warning performance includes response time and accuracy. The model has functions for structure initialization, parameter optimization, forward propagation, backward update, and model saving and import. S3: Online Identification and Interaction of Presentation Layer: The trained optimal model is integrated into the LabVIEW platform through the extern "C" interface to build a visual interface that includes sensor signal display, early warning result display and parameter adjustment functions. Real-time data transmission and inference are realized through upper and lower computer network communication to identify and warn of compressor stall surge conditions online.

2. The online identification method for stall and surge of aero-engine compressor based on deep learning according to claim 1, characterized in that, The specific steps of S1 include: S11: Real-time acquisition of compressor dynamic pressure signals. The acquisition location is on the cross-section of the compressor rotor leading edge. Pressure sensors are evenly arranged along the circumference. Data from two symmetrical measurement points out of six measurement points are selected. The sensor sampling frequency is 25600Hz. S12: Preprocess the acquired raw pressure signal, decompose the signal sequence using a recursive method, transform the even-numbered and odd-numbered parts separately, and complete the butterfly merging through a rotation factor to obtain the spectrum of the time-domain signal; perform a fast Fourier transform on the real-number input signal, extract the first half of the spectrum, calculate the normalized feature of log-magnitude on the spectrum, with the normalization standard being zero mean and unit standard deviation, and return a 512-dimensional normalized feature vector; S13: Perform time window segmentation on the preprocessed signal. The segmented data is used for subsequent model training and online identification.

3. The online stall and surge identification method for aero-engine compressors based on deep learning according to claim 1, characterized in that, The CNN model construction in step S2 includes: The input layer receives a 512-dimensional feature vector; the convolutional layer uses 8 convolutional kernels of size 5×1, with ReLU activation function and initial weights uniformly distributed in the range [-1,1]; the pooling layer uses max pooling with a window size of 2×1, preserving the location information of the maximum value for backpropagation; the fully connected layer calculates the loss by combining cross-entropy with the Softmax function. During backpropagation, the fully connected layer directly calculates the weight gradient, while the pooling layer only propagates the gradient to the location of the maximum value.

4. The online stall and surge identification method for aero-engine compressors based on deep learning according to claim 1, characterized in that, The LSTM model construction in step S2 includes: It contains one LSTM unit and one fully connected output layer. The LSTM unit consists of a forget gate, an input gate, an output gate, and candidate memories. The input dimension is 512, the hidden layer size is 16, the learning rate is 0.0001, the binary classification output dimension is 2, and the batch size is 32. The weights and biases are initialized with a uniform random distribution in the range of [-0.1, 0.1]. During forward propagation, the gate value, memory state, and hidden state of each time step are saved to the buffer for backpropagation. Finally, the output is calculated through the hidden state and the probability distribution is obtained by the Softmax function.

5. The online stall and surge identification method for aero-engine compressors based on deep learning according to claim 1, characterized in that, The CNN-GRU model construction in step S2 includes: It consists of a CNN feature extraction module and a GRU sequence modeling module connected in series. The CNN module takes 16-channel features as input, extracts features through convolutional layers, and outputs an 8-channel feature vector. The convolutional kernels are 5×1 in size and number 8. The GRU module receives the temporal feature vector output by the CNN, has 16 hidden layer neurons, captures long-term dependencies in the sequence through update and reset gates, and finally outputs the classification result through a fully connected layer.

6. The online identification method for stall and surge of aero-engine compressor based on deep learning according to claim 1, characterized in that, The specific steps of model training in step S2 include: S21: Dataset Partitioning: The preprocessed 512-dimensional feature vectors are divided into training and test sets in a 7:3 ratio. The training set is then further divided into sub-training and validation sets in a 4:1 ratio. S22: Model Training: Three models are constructed, all using mini-batch gradient descent to optimize parameters. The loss function for CNN and CNN-GRU is cross-entropy + Softmax, while the loss function for LSTM is cross-entropy. During training, the sub-training set loss is calculated in real-time and the weights are updated in reverse. After each training round, accuracy, precision, and response time are evaluated on the validation set. Training stops when there is no improvement in validation set performance for 5 consecutive rounds. The model parameters with the best performance on the validation set are saved as binary files. S23: Model Comparison: Under three operating conditions of 1000rpm, 1200rpm, and 1400rpm, the warning response time and accuracy of the three models were tested respectively, and the model with the best overall performance was selected as the main model for online identification.

7. The online stall and surge identification method for aero-engine compressors based on deep learning according to claim 1, characterized in that, The specific steps of S3 include: S31: Constructing an interactive system between the upper and lower computers: The upper computer deploys a LabVIEW development environment, including a visual interface. The visual interface includes three canvas areas displaying sensor 1 signals, sensor 2 signals, and early warning results, as well as four interactive modules: sensor coefficient calibration, threshold selection, early warning indicator, and a stop operation button. The lower computer uses an NI cRIO-9033 controller, integrating an FPGA module for high-speed signal acquisition, and a real-time processor loading cross-compiled model dynamic link libraries and parameter files. S32: Real-time identification process: The lower-level machine collects pressure signals at 25600Hz through the FPGA module, generates 512-dimensional feature vectors after preprocessing, and inputs them into the pre-trained CNN-GRU model. The model outputs stall and surge probability values. When the probability value exceeds the preset threshold of 0.85, the lower-level machine sends an early warning signal to the upper-level machine through the TCP / IP protocol. The upper-level machine interface triggers the early warning indicator and records the early warning time. S33: Model Deployment Adaptation: Deep learning models expose exported functions through the extern "C" interface. After cross-compiling with the Xilinx compiler, they are adapted to the embedded environment, supporting real-time inference latency of ≤50ms on the NI cRIO-9033 controller.

8. The online identification method for stall and surge of aero-engine compressor based on deep learning according to claim 1, characterized in that, The arrangement of the pressure sensors in step S1 satisfies the following conditions: they are distributed at 45° intervals along the circumference of the compressor rotor leading edge section, the sensor measurement range covers the full operating pressure value within the compressor surge boundary, and they are fixed to the steel frame structure platform by a bracket, with an axial clearance of 33mm between them and the compressor test piece.

9. The online identification method for stall and surge of aero-engine compressor based on deep learning according to claim 1, characterized in that, The parameters for the mini-batch gradient descent method used in step S2 are: momentum coefficient 0.9, weight decay coefficient 1e-5, and initial learning rate. 0.001, the learning rate is automatically decayed when the loss on the validation set stops decreasing, with a decay factor of 0.

5.

10. The online identification method for stall and surge of aero-engine compressor based on deep learning according to claim 1, characterized in that, In step S3, the communication between the upper and lower computers adopts the NI-RIO protocol, the data transmission format is JSON, which includes 5 fields: sensor ID, sampling timestamp, original pressure value, preprocessed feature vector, and model output probability. The transmission rate is ≥1Mbps and the packet loss rate is ≤0.1%.

Images