Aero-engine gas path fault diagnosis method combining continuous learning and neural network

Abstract

The method for aero-engine gas path fault diagnosis combining continuous learning and neural network belongs to the field of fault diagnosis. It includes the construction of aero-engine gas path multi-source sensor dataset; the design of a one-dimensional convolutional neural network-based backbone feature extraction network; the offline joint training of the backbone feature extraction network and the ladder dimension cosine prototype classifier PDP; the design of a network expansion mechanism for continuous learning training; the updating of the gas path fault database and model fine-tuning; and knowledge distillation. The lightweight student network Aero-KDNet designed based on knowledge distillation in the patent can realize the complex diagnostic capability of the teacher network with a small parameter size. Its efficient computing efficiency and extremely low memory occupation effectively overcome the computing power limitation of airborne equipment, laying a solid technical foundation for subsequent hardware device embedded deployment and the development of airborne online high-precision gas path fault monitoring.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis and relates to a method for diagnosing air circuit faults in aero-engines that combines continuous learning and neural networks. Background Technology

[0002] As the core power component of an aircraft, the aero-engine is also a crucial element of aircraft health management technology. With the development of China's aviation industry, the demands for intelligent aero-engines are increasing, placing higher requirements on engine airflow health detection solutions. Researching long-term, accurate airflow fault diagnosis technologies is essential for ensuring flight safety and preventing catastrophic accidents.

[0003] With the rapid development of artificial intelligence technology, deep learning methods have achieved remarkable success in various fields, especially demonstrating outstanding performance on static datasets. However, real-world application scenarios are often constantly changing, particularly when dealing with tasks that continuously acquire new data. Monitoring systems must adapt to these data changes while retaining previously learned knowledge. Traditional deep learning methods are typically trained on fixed datasets and assume that the training data is static. However, it is impossible to collect data on all fault categories during the test run of an aero-engine. When the model is deployed and faces dynamically changing task data, traditional deep learning methods often forget knowledge from historical data when processing new data, a problem known as catastrophic forgetting. To address catastrophic forgetting in dynamic task environments, continuous learning, by retaining historical knowledge and continuously learning new knowledge, can reduce the loss of knowledge from old tasks when learning new tasks, while ensuring the model's performance in new tasks.

[0004] As aero-engine structures become increasingly complex and the types of collected data expand, related intelligent methods are also becoming increasingly large and complex. This not only increases the computational burden in resource-constrained airborne computing scenarios but also reduces the processing performance of intelligent methods. To address the problem of redundant model parameters, model lightweighting has emerged, mainly including knowledge distillation, quantization, and pruning. Among these, knowledge distillation, as an important technique that balances model compression and accuracy preservation, has attracted researchers' attention. For example, Yuxiang Kang et al., in their paper "Fault anomaly detection method of aero-engine rollingbearing based on distillation learning," utilized knowledge distillation and Vision Transformer to construct a fault diagnosis method for aero-engine bearings with fewer parameters. Xiangkun Sun et al., in their paper "SDD-DETR: SurfaceDefect Detection for No-Service Aero-Engine Blades With DetectionTransformer," used a lightweight Transformer network to complete the detection of surface defects on aero-engine blades. Fault diagnosis methods based on continuous learning, knowledge distillation, and neural networks have achieved research results in many fields, but research on such methods in the problem of aero-engine gas path fault diagnosis remains limited. Summary of the Invention

[0005] To overcome the limitations of neural networks in learning from newly added air path fault categories, and to reduce the number of parameters in neural network models and improve hardware compatibility through knowledge distillation, this invention proposes a method for diagnosing air path faults in aero-engines that combines continuous learning with neural networks. This method, through the design of the Aero-CLNet network, achieves the simultaneous learning of new air path fault categories and the retention of knowledge about historical air path fault categories. The knowledge distillation scheme ensures that when the Aero-CLNet network encounters new air path fault category data, the core feature extraction network is frozen and lightweight extension network branches are dynamically added. This avoids the computational speed decrease caused by increased network parameters. The student model Aero-KDNet obtained through knowledge distillation effectively reduces the number of parameters while maintaining fault diagnosis performance.

[0006] The technical solution of the present invention: The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks specifically includes the following steps: S1. Construct a multi-source sensor dataset for the air path of an aero-engine; First, historical operating data of the aero-engine's air path is acquired, and a multi-source sensor dataset of the air path is constructed. The historical operating data of the air path includes time series data from the multi-source sensor data of the air path. and corresponding gas path fault category labels .

[0007] S2. Design a backbone feature extraction network based on a one-dimensional convolutional neural network. ; The backbone feature extraction network is used to extract temporal information features and spatial coupling features from multi-source sensor data in the gas path. The specific process is as follows: Let the time series of the input gas path multi-source sensor data be... ,in Number of sensor channels The time series length is specified. The backbone feature extraction network consists of several stacked basic neural network modules, each processing the input gas path multi-source sensor data. 1) 1D-CNN layer: Extracts temporal information features using sliding convolution kernels, the first... l Layer network j The output channel is in the first t The convolution output feature values ​​at each time step are represented as: in, l Indicates the network layer number; j Indicates the output channel number; i This indicates the input channel number, which corresponds to the sensor channel number. t Indicates the time step number; K This represents the length of the one-dimensional convolution kernel. k This represents the position index inside a one-dimensional convolution kernel; The sliding stride is the size of the convolution operation. Indicates the first l -1 layer network i The input channel in the first tS + k Feature values ​​at each position; Indicates the first l Connecting the first layer in the network i The input channel and the first j A one-dimensional convolutional kernel with output channels in the th... k Weights at each position; Indicates the first l Layer network j The bias parameters corresponding to each output channel.

[0008] 2) Batch Normalization Layer: Used to eliminate the offset of convolutional output feature values ​​in 1D-CNN layers, eliminate the magnitude difference between different sensors, and accelerate the convergence of the backbone feature extraction network. It performs batch normalization on the convolutional output feature values ​​of the 1D-CNN layers. in, and Learnable scaling and translation affine parameters for batch normalization layers; To prevent tiny constants with a denominator of zero; B This represents the current mini-batch of samples; Indicates the first l Layer network j Each output channel at time step t Batch normalized output eigenvalues; and These are the mean and variance of the current mini-batch samples, respectively.

[0009] 3) Nonlinear Activation Layer (ReLU): Introduces nonlinear mapping capability, enabling the backbone feature extraction network to approximate high-dimensional and complex fault boundaries. in, Indicates the first l Layer network j Each output channel at time step t The nonlinear activation output eigenvalues; This indicates the operation of retrieving the maximum value; 4) MaxPool1d: While preserving the most significant fault information from the multi-source sensor data in the gas path, it performs dimensionality reduction to improve the translation invariance of the nonlinear activation output eigenvalues. in, Indicates the first l Layer network j The output channel is in the first t The maximum pooling result at each time step; Indicates the first l Layer network j The output channel is in the first t The time step, the first q The time dimension position number covered by each pooling window The nonlinear activation output eigenvalue at the location; This is the pooling window size; This is the pooling step size.

[0010] 5) Random Dropout Layer: To prevent the backbone feature extraction network from overfitting to a limited number of faulty samples during the training phase, the outputs of some neurons are randomly set to zero according to a given dropout probability. in, Indicates the first l Layer network j Each output channel at time step t The feature values ​​obtained after random inactivation processing; Indicates the first l Layer network j Each output channel at time step t The corresponding binary random mask.

[0011] After several basic neural network modules are stacked and computed, a three-dimensional feature map is output. The three-dimensional feature map will be flattened by a one-dimensional fully connected layer, and the three-dimensional feature map will be compressed into a basic high-dimensional feature vector with fixed dimensions.

[0012] S3. Offline joint training of backbone feature extraction network and step dimension cosine prototype classifier PDP; Each time series sample in the gas path multi-source sensor dataset Input the main feature extraction network sequentially For each time series sample Each of them yields its corresponding fixed-dimensional basic high-dimensional feature vector: Among them, superscript This indicates the first stage, namely the offline joint training stage; This represents the network parameters of the backbone feature extraction network; This indicates the size of a fixed dimension.

[0013] Simultaneously, the centroid vectors of the category features corresponding to various gas path faults are initialized in the classification layer of the PDP. The set of centroid vectors of all historical categorical features Network parameters of the backbone feature extraction network Meanwhile, during the offline joint training phase, it is optimized through backpropagation as a learnable parameter.

[0014] To achieve end-to-end offline joint training, PDP will use the basic high-dimensional feature vectors With the centroid vector of category features Mapping to a norm metric space. Specifically: for any time series sample of gas path multi-source sensor data. First, the backbone feature extraction network obtains its basic high-dimensional feature vector. Then the PDP calculates the scaling factor. The cosine similarity is used to output the label of the historical gas path fault category of the time series sample. c Posterior probability: in, This represents the first category in the set of historical gas path fault categories covered during the current offline joint training phase. The centroid vector of the category features corresponding to each gas circuit fault category; This represents the set of gas path fault categories used to traverse the history in the denominator of the formula. Index of candidate gas path fault categories; Time series samples The actual gas path fault category labels; cosine similarity is defined as... L Dot product normalized to 2-norm: , Represents the basic high-dimensional feature vector The transpose of ; Represents the basic high-dimensional feature vector of L 2-norm and centroid vector of categorical features of L 2. Product of norms.

[0015] The parameters of the backbone feature extraction network and the centroid vector of the category features are optimized by minimizing the cosine cross-entropy loss function with a scaling factor, ensuring the centroid vector of the category features is optimized. It can maximize the cosine similarity between time series samples of the same class and minimize the cosine similarity between time series samples of different classes. (Single time series sample) The cosine cross-entropy loss function is defined as: The posterior probability formula is used to calculate a single time series sample. The probability distribution of each candidate gas path fault category. During the offline joint training phase, time series samples... Real gas path fault category tags Since this is known, a cosine cross-entropy loss function with a scaling factor is constructed using the posterior probability corresponding to the actual gas path fault category label. By minimizing the cosine cross-entropy loss function, the underlying high-dimensional feature vector corresponding to the time series samples is optimized. Its true class feature centroid vector The cosine similarity increases, and at the same time, relative to the centroid vectors of other categories of features, the similarity increases. The cosine similarity is reduced, thus completing the joint optimization of the network parameters of the backbone feature extraction network and the centroid vector of the category features. Once offline joint training is completed, the network parameters of the backbone feature extraction network will be frozen. And the set of centroid vectors of category features corresponding to all historical gas path fault categories. And forcibly lock the final statistical mean and variance after offline joint training of the batch normalized layers. This allows for the preservation of historical knowledge.

[0016] like Figure 2 As shown, the Aero-CLNet network is designed, which includes a backbone feature extraction network. With lightweight extended network branch Therefore, the training of the Aero-CLNet network includes offline joint training of the main feature extraction network based on historical gas path operation data and continuous learning training of the lightweight extended network branches based on data of newly added gas path fault categories. The network parameters of the frozen backbone feature extraction network were obtained through offline joint training in step S3. Historical gas circuit fault categories The set of centroid vectors of historical category features And the statistical mean and variance of the batch normalized layer The basic high-dimensional feature vector is calculated by judging the input of new time series samples into the backbone feature extraction network. Centroid vector of historical category features If the maximum cosine similarity is less than the threshold, the new time series sample is determined not to belong to the historical gas path fault category, is identified as a newly added gas path fault category, and the continuous learning and training of the lightweight extended network branch in step S4 is triggered.

[0017] S4. Design a network expansion mechanism for continuous learning and training; Lightweight extended network branch This is a parallel network structure. The lightweight extended network branch includes a single-layer 1D-CNN layer, a single-layer batch normalization layer, a single-layer nonlinear activation layer, and a single-layer global average pooling layer. The single-layer 1D-CNN layer, batch normalization layer, and nonlinear activation layer are the same as those in the backbone feature extraction network constructed by S2, and the size of the one-dimensional convolutional kernel in the 1D-CNN layer is 1. By setting a smaller one-dimensional convolutional kernel and a global average pooling layer (GAP), the parameters of the lightweight extended network branch are lower than those of the backbone feature extraction network. The extended high-dimensional feature vector is obtained after processing by the global average pooling layer, expressed as: in, This represents the total length of the features input into GAP in the time dimension. Indicates the first s The first branch of lightweight scaling networks in phase expansion j Each output channel at time step t The local features of the output of the nonlinear activation layer.

[0018] When faced with a new type of gas circuit fault, a continuous learning training session is conducted for each new type of gas circuit fault. The continuous learning training session is conducted on the [number]th ... s The stage corresponds to the newly added first s This type of gas circuit malfunction corresponds to the first... s A lightweight extension network branch is added; at this time, the backbone feature extraction network keeps its network parameters frozen, and the newly added lightweight extension network branch is activated. Feature extraction is performed. Lightweight extended network branch. New time series samples will be output. Extended high-dimensional feature vectors , For the first s Network parameters corresponding to each lightweight extended network branch The dimension of the extended high-dimensional feature vector extracted for the lightweight extended network branch.

[0019] New time series samples during inference in diagnostic tasks The global concatenation feature vector is output through the backbone feature extraction network and the lightweight extension network branches. Depending on whether the gas path fault category is known, the basic high-dimensional feature vector or the globally concatenated feature vector will be used by the PDP to calculate the Logit score for each category using a step-wise dimensionality truncation matching mechanism. The Logit score will then be backpropagated through a cosine cross-entropy loss function with a scaling factor to train the newly added lightweight extended network branch and the newly added category feature centroid vector. At the same time, the Logit score is also used for gas path fault category discrimination during the inference stage.

[0020] The step-by-step dimensional truncation matching mechanism is as follows: For historical gas circuit fault categories The globally concatenated feature vector is truncated, and the Logit score is calculated using only the basic high-dimensional feature vector output by the backbone feature extraction network. For the newly added gas circuit fault categories By using globally concatenated feature vectors and newly added class feature centroid vectors Calculate the Logit score: in, Indicates the first s A lightweight extended network branch; This represents the newly added set of gas path fault categories. In the... s After each stage is completed, the newly added gas path fault categories will be incorporated into the known gas path fault category set, and the newly learned category feature centroid vectors will be added to the category feature centroid vector set for use in subsequent stages for gas path fault category discrimination and historical knowledge retention.

[0021] During the continuous learning process, since the network parameters of the backbone feature extraction network are frozen, the network optimizer only updates the network parameters of the extended lightweight extension network branches. With the newly added category feature centroid vector By leveraging this dynamic structural expansion and similarity-based selection mechanism, Aero-CLNet achieves adaptive adjustment to new faults without disrupting the historical gas path fault decision boundaries, effectively addressing the catastrophic forgetting problem. This mechanism ensures the model's continuous learning capability while adding very few model parameters.

[0022] S5. Update the gas path fault database and fine-tune the model; Step S4 yields the newly identified gas path fault categories and their corresponding time series samples, the newly added category feature centroid vectors, and the network parameters corresponding to the lightweight extended branches.

[0023] Based on the newly identified gas path fault categories and corresponding time series samples determined in step S4, the gas path multi-source sensor dataset constructed in S1 is updated. A playback pool is built for use with the data added in S4. s Fine-tuning of the model for each lightweight extended network branch. The specific process of model fine-tuning is as follows: The time series samples in the replay pool... Input Aero-CLNet to obtain global concatenated features, calculate the Logit score of time series samples and centroid vectors of all class features, and fine-tune the network parameters of newly added class feature centroid vectors and lightweight extended network branches using the cross-entropy loss function.

[0024] The replay pool simultaneously introduces time-series samples of newly added gas path fault categories and historical time-series samples. The time-series samples of the newly added gas path fault categories enable the newly added lightweight extended network branches to represent the features of the new gas path faults. The time-series samples of historical gas path fault categories impose constraints on the output of the newly added lightweight extended network branches, reducing the interference of the newly added lightweight extended network branches on the feature space and classification boundaries of historical gas path fault categories, and mitigating the class bias and catastrophic forgetting caused by training only with time-series samples of the newly added gas path fault categories. The time-series samples of the newly added gas path fault categories and the time-series samples of historical gas path fault categories simultaneously optimize the centroid vector of the new category features, making the globally concatenated feature vector corresponding to the newly added time-series samples close to its corresponding centroid and far away from other category centroids. Representative samples of historical gas path fault categories are selected in the replay pool based on the cosine similarity between their basic high-dimensional feature vectors and the corresponding category feature centroid vectors. For the scarce data of newly added gas path fault categories, oversampling and data jitter injection are used to enhance the model and prevent it from overfitting.

[0025] To avoid interference from coupled gas path faults in the classification results, this patent uses a PDP to replace the traditional Softmax fully connected classification layer. L 2-norm normalization calculates the cosine similarity between the globally concatenated feature vector and the centroid vector of the category features within the corresponding norm metric space. This mechanism not only solves the catastrophic forgetting caused by model parameter inflation but also accurately identifies the feature differences between coupled gas path faults and single historical categories, thus improving the identification accuracy of coupled gas path faults.

[0026] S6. Knowledge distillation; As diagnostic tasks continue, more and more new gas path fault categories will be added, making the lightweight extension branches of Aero-CLNet increasingly complex and redundant. To adapt to computationally limited scenarios, knowledge distillation is used to lightweight the Aero-CLNet network. The Aero-CLNet network fine-tuned in step S5 is used as the teacher network, and the time-series samples in the replay pool constructed in step S5 are input into the teacher network to obtain the soft label distribution of the teacher network output. The time-series samples in the same replay pool are input into the student network Aero-KDNet, and the output distribution of the student network is constrained to approximate the output distribution of the teacher network through KL divergence loss, finally obtaining the lightweight gas path fault diagnosis model Aero-KDNet for airborne deployment. The student network Aero-KDNet obtained in this step has the same set of gas path fault category label outputs as the teacher network Aero-CLNet. The network structure of Aero-KDNet includes a single-layer 1D-CNN layer, a single-layer nonlinear activation layer, a single-layer max-pooling layer, a single-layer flattening layer, and a single-layer fully connected layer.

[0027] The beneficial effects of this invention are as follows: Through a two-layer architecture combining continuous learning and knowledge distillation techniques, an innovative method for diagnosing aero-engine gas path faults is proposed, integrating continuous learning and neural networks. This method combines the advantages of continuous learning in dealing with sudden unknown gas path faults and mitigating catastrophic forgetting with the characteristics of knowledge distillation in network model parameter compression and potential knowledge extraction, effectively improving the classification and identification accuracy of gas path fault diagnosis methods when facing unknown and coupled gas path faults. Simultaneously, relying on the feature constraints of a stepped-dimensional cosine prototype classifier, it ensures that the neural network possesses feature memory and global feature vector decoupling capabilities. More importantly, this scheme effectively alleviates the computational waste problem of traditional deep learning models requiring retraining in data accumulation environments, providing a technical path reference for the application of artificial intelligence technology in the full-cycle health management of aero-engines. The lightweight student network Aero-KDNet, designed based on knowledge distillation, can achieve the complex diagnostic capabilities of a teacher network with a relatively small parameter scale. Its high computational efficiency and extremely low memory usage effectively overcome the computational limitations of airborne equipment, laying a solid technical foundation for subsequent embedded deployment of hardware and the implementation of high-precision online air path fault monitoring. Attached Figure Description

[0028] Figure 1 A flowchart of a gas path fault diagnosis scheme based on a two-layer technical architecture for continuous learning and knowledge distillation.

[0029] Figure 2 The diagram shows the structure of the Aero-CLNet and Aero-KDNet neural networks.

[0030] Figure 3 The feature visualization diagrams obtained by t-SNE for gas circuit fault diagnosis in stages 1, 2, and 3 are shown. Detailed Implementation

[0031] The present invention will be further described below with reference to the technical solution and accompanying drawings.

[0032] The flowchart of the aero-engine air path fault diagnosis method combining continuous learning and neural networks is as follows: Figure 1 As shown. Simulation verification is as follows: 1) Data preprocessing; This invention will use a typical air path fault scenario of an aero-engine air path system to verify the effectiveness of the proposed air path fault diagnosis method. The collected data includes time series data from multiple air path sensors. and corresponding gas path fault labels The detailed information of the data is shown in Table 1.

[0033] Table 1: Names and Meanings of Data Variables from Multi-Source Sensors in the Gas Path

[0034] First, the collected gas path multi-source sensor data is normalized, as follows: Because the data from various sensors in the aero-engine control system vary greatly in magnitude, all the data involved in this patent need to be normalized. in, This represents the time series of normalized sensor data. x This represents the time series of sensor data before normalization. This represents the maximum value of the time series data from this sensor within the sample set. This represents the minimum value of the time series data from this sensor within the sample set.

[0035] 2) Construct continuous learning tasks; Gas path failures mainly consist of single gas path failures of gas path components: fan, low-pressure compressor (LPC), high-pressure compressor (HPC), high-pressure turbine (HPT), and low-pressure turbine (LPT), as well as coupled gas path failures involving multiple components. Common gas path failure categories, causes, and characteristics are shown in Table 2. Categories C0 to C2 were selected as the initial gas path failure dataset. Categories C3 and C4 were added as two new gas path failure categories for simulation verification, with the sample size of C3 and C4 being much smaller than that of C0 to C2 to simulate an imbalanced sample ratio of new gas path failures. Gas path failure C3 differs from gas path failure categories C0 to C2, while gas path failure C4 can be considered a coupled form of C2 and C3 gas path failures. Considering different gas path failure verification scenarios can comprehensively verify the effectiveness of the proposed gas path failure diagnosis scheme.

[0036] Table 2: Common Gas Circuit Fault Categories, Causes, and Characteristics

[0037] 3) Network training parameter settings; In this patent, the dataset required for neural network training is generated from an aero-engine airflow system model. Airflow faults are simulated by adjusting the health factors of corresponding components. The sampling time is 0.04 seconds. For historical airflow fault samples, 5000 data points are collected to construct a sample set of length 50 and quantity 1000. For newly added airflow fault samples, 1000 data points are collected to construct a dataset of length 50 and quantity 200 to simulate the sparse airflow fault data in real-world scenarios. The training and test sets are divided in an 8:2 ratio, and the training batch size is set to 32.

[0038] The model training and distillation processes were implemented using the PyTorch deep learning framework. During iteration, the number of epochs for both offline joint training and continuous learning was set to 30, while the number of epochs for knowledge distillation was set to 15. The initial learning rate was set to 5e-3, and an Adam network optimizer with weight decay was employed. Both the training and validation processes of the neural network were accelerated using an NVIDIA GPU RTX 3090.

[0039] 4) Analysis of model test results; To quantitatively evaluate the combined ability of neural networks to "learn new knowledge" and "retain old knowledge" when faced with unknown gas path faults, incremental accuracy was selected. and incremental forgetting rate As an evaluation indicator. This reflects the importance of gas circuit fault diagnosis methods after learning the first step. s The average identification accuracy of all encountered gas path fault categories after each stage of the task represents the comprehensive diagnostic performance of the gas path fault diagnosis method. This measures the average decrease in accuracy of gas path fault diagnosis methods for identifying historical tasks after learning a new task, intuitively reflecting the ability of gas path fault diagnosis methods to resist "catastrophic forgetting." The calculation formula is as follows: in This indicates that the gas circuit fault diagnosis method was just learned after completing the task. Immediately during the mission The accuracy rate obtained from the above test. This involves learning the gas path fault diagnosis method after completing the subsequent tasks. Afterwards, the task was tested again. The obtained accuracy rate.

[0040] Table 3: Comparison of performance metrics of test datasets after distillation for different neural networks

[0041] Table 3 compares the performance metrics of different neural networks on test datasets at various stages of continuous learning. Figure 3As shown, a comparative experiment was conducted by replacing the 1DCNN-PDP using PDP within Aero-KDNet with the 1DCNN-FC fully connected classifier. It can be seen that 1DCNN-PDP performs the same as or better than 1DCNN-FC in all three continuous learning stages in terms of evaluation metrics. This indicates that the lightweight extended network branch and PDP can improve the accuracy of gas path fault classification under dynamic and coupled conditions. Simultaneously, the lightweight student network obtained through knowledge distillation retains diagnostic capabilities with a very small parameter scale, meeting the stringent memory and computing power constraints of airborne edge devices.

Claims

1. A method for diagnosing aero-engine air path faults by combining continuous learning and neural networks, characterized in that: Specifically, it includes the following steps: S1. Construct a multi-source sensor dataset for the air path of an aero-engine; S2. Design a backbone feature extraction network based on a one-dimensional convolutional neural network. ; S3. Offline joint training of backbone feature extraction network and step dimension cosine prototype classifier PDP; S4. Design a network expansion mechanism for continuous learning and training; S5. Update the gas path fault database and fine-tune the model; S6. Knowledge distillation.

2. The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks according to claim 1, characterized in that, Step S1 specifically involves: firstly acquiring historical operating data of the aero-engine's air path, and constructing a multi-source sensor dataset for the air path. The historical operating data of the air path includes the time series of multi-source sensor data. and corresponding gas path fault category labels .

3. The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks according to claim 2, characterized in that, Step S2 specifically involves the following: The backbone feature extraction network is used to extract the temporal information features and spatial coupling features of the multi-source sensor data in the gas path, as follows: Let the time series of the input gas path multi-source sensor data be... ,in Number of sensor channels The time series length is specified; the backbone feature extraction network consists of several stacked basic neural network modules, each of which processes the input gas path multi-source sensor data. 1) 1D-CNN layer: Extracts temporal information features using sliding convolution kernels, the first... l Layer network j The output channel is in the first t The convolutional output feature values ​​at each time step are represented as: in, l Indicates the network layer number; j Indicates the output channel number; i This indicates the input channel number, which corresponds to the sensor channel number. t Indicates the time step number; K This represents the length of the one-dimensional convolution kernel. k This represents the position index inside a one-dimensional convolution kernel; The stride is the sliding step size for the convolution operation; Indicates the first l -1 layer network i The input channel in the first tS + k Feature values ​​at each position; Indicates the first l Connecting the first layer in the network i The input channel and the first j A one-dimensional convolutional kernel with output channels in the th... k Weights at each position; Indicates the first l Layer network j The bias parameters corresponding to each output channel; 2) Batch Normalization Layer: Performs batch normalization on the convolutional output feature values ​​of the 1D-CNN layer. in, and Learnable scaling and translation affine parameters for batch normalization layers; To prevent tiny constants with a denominator of zero; B This represents the current mini-batch of samples; Indicates the first l Layer network j Each output channel at time step t Batch normalized output eigenvalues; and These are the mean and variance of the current mini-batch samples, respectively. 3) Nonlinear activation layer: Introducing nonlinear mapping capability enables the backbone feature extraction network to approximate high-dimensional and complex fault boundaries. in, Indicates the first l Layer network j Each output channel at time step t The nonlinear activation output eigenvalues; This indicates the operation of retrieving the maximum value; 4) Max Pooling Layer: While retaining the most significant fault information from the multi-source sensor data in the gas path, dimensionality reduction is performed to improve the translation invariance of the nonlinear activation output eigenvalues. in, Indicates the first l Layer network j The output channel is in the first t The maximum pooling result at each time step; Indicates the first l Layer network j The output channel is in the first t The time step, the first q The time dimension position number covered by each pooling window The nonlinear activation output eigenvalue at the location; This is the pooling window size; The pooling step size; 5) Random deactivation layer: Randomly set the output of some neurons to zero according to a given discard probability. in, Indicates the first l Layer network j Each output channel at time step t The feature values ​​obtained after random inactivation processing; Indicates the first l Layer network j Each output channel at time step t The corresponding binary random mask; After several basic neural network modules are stacked and computed, a three-dimensional feature map is output. The three-dimensional feature map will be flattened by a one-dimensional fully connected layer, and the three-dimensional feature map will be compressed into a basic high-dimensional feature vector with fixed dimensions.

4. The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks according to claim 3, characterized in that, Step S3 specifically involves: processing each time series sample in the gas path multi-source sensor dataset. Input the main feature extraction network sequentially For each time series sample Each of them yields its corresponding fixed-dimensional basic high-dimensional feature vector: Among them, superscript This indicates the first stage, namely the offline joint training stage; This represents the network parameters of the backbone feature extraction network; Indicates the size of a fixed dimension; Simultaneously, the centroid vectors of the category features corresponding to various gas path faults are initialized in the classification layer of the PDP. The set of centroid vectors of all historical categorical features Network parameters of the backbone feature extraction network Simultaneously, during the offline joint training phase, it is optimized through backpropagation as a learnable parameter. To achieve end-to-end offline joint training, PDP will use the basic high-dimensional feature vectors With the centroid vector of category features Mapping to a norm metric space; specifically: for any time series sample of gas path multi-source sensor data. First, the backbone feature extraction network obtains its basic high-dimensional feature vector. Then the PDP calculates the scaling factor. The cosine similarity is used to output the label of the historical gas path fault category of the time series sample. c Posterior probability: in, This represents the first category in the set of historical gas path fault categories covered during the current offline joint training phase. The centroid vector of the category features corresponding to each gas circuit fault category; This represents the set of gas path fault categories used to traverse the history in the denominator of the formula. Index of candidate gas path fault categories; Time series samples The actual gas path fault category labels; cosine similarity is defined as... L The dot product after 2-norm normalization: , Represents the basic high-dimensional feature vector The transpose of ; Represents the basic high-dimensional feature vector of L 2-norm and centroid vector of categorical features of L 2-norm product; The parameters of the backbone feature extraction network and the centroid vector of the category features are optimized by minimizing the cosine cross-entropy loss function with a scaling factor, ensuring the centroid vector of the category features is optimized. It can maximize the cosine similarity between time series samples of the same class and minimize the cosine similarity between time series samples of different classes; a single time series sample The cosine cross-entropy loss function is defined as: The posterior probability formula is used to calculate a single time series sample. The probability distribution of each candidate gas path fault category; by minimizing the cosine cross-entropy loss function, the basic high-dimensional feature vector corresponding to the time series samples is obtained. Its true class feature centroid vector The cosine similarity increases, and at the same time, relative to the centroid vectors of other categories of features, the similarity increases. The cosine similarity is reduced, thus completing the joint optimization of the network parameters of the backbone feature extraction network and the centroid vector of the category features; once offline joint training is completed, the network parameters of the backbone feature extraction network will be frozen. And the set of centroid vectors of category features corresponding to all historical gas path fault categories. And forcibly lock the final statistical mean and variance after offline joint training of the batch normalized layers. This is to preserve historical knowledge; Design the Aero-CLNet network, which includes a backbone feature extraction network. With lightweight extended network branch Therefore, the training of the Aero-CLNet network includes offline joint training of the main feature extraction network based on historical gas path operation data and continuous learning training of the lightweight extended network branches based on data of new gas path fault categories. The network parameters of the frozen backbone feature extraction network were obtained through offline joint training in step S3. Historical gas circuit fault categories The set of centroid vectors of historical category features And the statistical mean and variance of the batch normalized layer The basic high-dimensional feature vector is calculated by judging the input of new time series samples into the backbone feature extraction network. Centroid vector of historical category features If the maximum cosine similarity is less than the threshold, the new time series sample is determined not to belong to the historical gas path fault category, is identified as a newly added gas path fault category, and the continuous learning and training of the lightweight extended network branch in step S4 is triggered.

5. The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks according to claim 4, characterized in that, Step S4 specifically involves: lightweight expansion of the network branch. This is a parallel network structure; the lightweight extended network branch includes a single-layer 1D-CNN layer, a single-layer batch normalization layer, a single-layer nonlinear activation layer, and a single-layer global average pooling layer. The single-layer 1D-CNN layer, batch normalization layer, and nonlinear activation layer are the same as those in the backbone feature extraction network constructed by S2, and the size of the one-dimensional convolutional kernel in the 1D-CNN layer is 1. By setting a smaller one-dimensional convolutional kernel and a global average pooling layer (GAP), the parameters of the lightweight extended network branch are lower than those of the backbone feature extraction network. The extended high-dimensional feature vector is obtained after processing by the global average pooling layer (GAP), expressed as: in, This represents the total length of the features input into GAP in the time dimension. Indicates the first s The first branch of lightweight scaling networks in phase expansion j Each output channel at time step t Local features of the output of the nonlinear activation layer; When faced with a new type of gas circuit fault, a continuous learning training session is conducted for each new type of gas circuit fault. The continuous learning training session is conducted on the [number]th ... s The stage corresponds to the newly added first s This type of gas circuit malfunction corresponds to the first... s A lightweight extension network branch is added; at this time, the backbone feature extraction network keeps its network parameters frozen, and the newly added lightweight extension network branch is activated. Feature extraction; lightweight expansion of network branches New time series samples will be output. Extended high-dimensional feature vectors , For the first s Network parameters corresponding to each lightweight extended network branch The dimension of the extended high-dimensional feature vector extracted for the lightweight extended network branch; New time series samples during inference in diagnostic tasks The global concatenation feature vector is output through the backbone feature extraction network and the lightweight extension network branches. Depending on whether the gas path fault category is known, the basic high-dimensional feature vector or the global concatenated feature vector will be calculated by the PDP using a step-wise dimensionality truncation matching mechanism to obtain the Logit score for each category. The Logit score will then be backpropagated through a cosine cross-entropy loss function with a scaling factor to train the newly added lightweight extended network branch and the newly added category feature centroid vector. At the same time, the Logit score is also used for gas path fault category discrimination during the inference stage. The step-by-step dimensional truncation matching mechanism is as follows: For historical gas circuit fault categories The globally concatenated feature vector is truncated, and the Logit score is calculated using only the basic high-dimensional feature vector output by the backbone feature extraction network. For the newly added gas circuit fault categories By using globally concatenated feature vectors and newly added class feature centroid vectors Calculate the Logit score: in, Indicates the first s A lightweight extended network branch; This represents the newly added set of gas path fault categories; in the... s After each stage is completed, the newly added gas path fault categories will be incorporated into the known gas path fault category set, and the newly learned category feature centroid vectors will be added to the category feature centroid vector set for use in subsequent stages for gas path fault category discrimination and historical knowledge retention.

6. The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks according to claim 5, characterized in that, Specifically, step S5 involves obtaining the newly identified gas path fault categories and corresponding time series samples, the centroid vector of the newly identified category features, and the network parameters corresponding to the lightweight extended branch in step S4. Based on the newly added gas path fault categories and corresponding time series samples identified in step S4, update the gas path multi-source sensor dataset constructed in S1. Build a replay pool for use in S4. s Fine-tuning of the model for each lightweight extended network branch; the specific process of model fine-tuning is as follows: time series samples from the replay pool... Input Aero-CLNet to obtain global concatenated features, calculate the Logit score of time series samples and centroid vectors of all class features, and fine-tune the network parameters of newly added class feature centroid vectors and lightweight extended network branches using the cross-entropy loss function.

7. The method for diagnosing aero-engine air path faults by combining continuous learning and neural networks according to claim 6, characterized in that, Step S6 specifically involves: using knowledge distillation to lightweight the Aero-CLNet network; using the Aero-CLNet network fine-tuned in step S5 as the teacher network, and inputting the time-series samples in the replay pool constructed in step S5 into the teacher network to obtain the soft label distribution of the teacher network output; inputting the time-series samples in the same replay pool into the student network Aero-KDNet, and constraining the student network output distribution to approximate the teacher network output distribution through KL divergence loss, finally obtaining the lightweight gas path fault diagnosis model Aero-KDNet for airborne deployment; the student network Aero-KDNet obtained in this step has the same gas path fault category label output set as the teacher network Aero-CLNet, and the network structure of Aero-KDNet includes a single-layer 1D-CNN layer, a single-layer nonlinear activation layer, a single-layer max pooling layer, a single-layer flattening layer, and a single-layer fully connected layer.

Images