Engine gas path fault diagnosis method based on domain adversarial neural network

Abstract

The present invention discloses a method for diagnosing engine gas path faults based on a domain adversarial neural network. In the actual engine data collection process, generally only a large amount of normal category data can be collected, and fault data is very small or even some fault category data is missing. In response to this problem, the present invention proposes a partially aligned domain adversarial neural network algorithm, which uses labeled data in the source domain and normal category data in the target domain to train a target domain diagnosis model, and accordingly designs a feature extractor, category classifier, and domain classifier. The present invention is used to identify various types of turbofan engine faults, including fan, compressor, high-pressure turbine, and low-pressure turbine faults. In the face of the situation where there is only normal data in the target domain, the domain adaptation method in transfer learning is used to utilize the knowledge contained in a large amount of data in the source domain to construct a fault diagnosis model with good performance. It provides a new method for diagnosing turbofan engine gas path faults in the absence of fault data.

Description

Technical Field

[0001] The present invention relates to the field of turbofan engine air path fault diagnosis, and in particular to an engine air path fault diagnosis method based on a domain adversarial neural network. Background Art

[0002] Aircraft engines are the power source of aircraft flight, and their safety is closely linked to the safety of the aircraft. An aircraft engine failure poses a significant threat to flight safety, resulting in significant financial losses at best and fatalities at worst. Therefore, failures must be avoided as much as possible during aircraft and engine operation.

[0003] However, aircraft engines operate in a harsh environment, subject to prolonged operation under conditions of high temperature, high pressure, high speed, and intense vibration, making them highly susceptible to failure. Over the life of an aircraft engine, component performance can degrade due to blade corrosion, erosion, fouling, wear, and increased tip clearance. Severe performance degradation can lead to failure. Furthermore, impact from foreign objects can cause a sharp drop in engine performance, leading to engine failure. Statistics show that 90% of aircraft engine failures are caused by failures of air path components. These components refer to several core components of an aircraft engine, including the fan, compressor, turbine, combustor, and exhaust nozzle. Maintenance costs for these components account for 60% of the total maintenance costs of an aircraft engine.

[0004] Currently, there are two common approaches to aircraft engine maintenance: scheduled maintenance and condition-based maintenance. Scheduled maintenance involves performing major or minor repairs on aircraft engines at fixed intervals, while condition-based maintenance monitors the engine's condition and repairs it if a fault occurs. Scheduled maintenance is easy to implement, but it can result in delayed repairs and wasteful labor and resources. Condition-based maintenance, however, resolves the issues associated with scheduled maintenance by determining whether to perform repairs based on the engine's condition. However, condition-based maintenance requires the establishment of a precise aircraft engine condition monitoring and fault diagnosis system.

[0005] Most research on data-driven aircraft engine fault diagnosis algorithms begins by collecting historical engine operating data, then dividing the data into training and validation sets. This data is then used to train a fault diagnosis model, and finally the engine's real-time operating data is input into the fault diagnosis model to obtain the engine fault diagnosis results. This approach assumes that the collected historical engine operating data and the current engine's real-time operating data belong to the same distribution. However, in real-world scenarios, due to differences in engine operating conditions, service time, and individual engine variations, there will be certain differences between the collected historical operating data and the current engine's operating data. Therefore, the engine fault diagnosis model trained with this data may not be suitable for the engine in its current state.

[0006] Transfer learning can address these issues. Its key idea is to transfer annotated data or knowledge structures from related fields to complete or improve learning outcomes in the target domain or task. This means collecting data from other engine operating conditions or states and transferring the knowledge structures related to engine health to the current engine operating condition or state, thereby helping to establish accurate fault diagnosis models using the limited data currently collected.

[0007] Transfer learning methods can be categorized into instance-based, feature-based, and parameter-based methods. Instance-based transfer learning methods primarily use instance weighting strategies, assigning greater weights to a subset of samples in the source domain so that the probability density distribution of these samples is similar to that of the target domain. Feature-based methods transform the original features to form new feature representations in a high-dimensional space, bringing the source and target domain data distributions closer under the new feature representations. Maximum mean discrepancy (MMD) is a commonly used distance metric. Parameter-based transfer learning methods transfer knowledge at the model or parameter level. This method assumes that model parameters represent the knowledge learned by the model from the data, and transferring model parameters is equivalent to transferring knowledge. These methods are often used in neural network-based models.

[0008] The domain adaptation method used in the present invention is a type of feature transformation method. The result of feature transformation is not measured by the "distance" between the two domains, but an adversarial neural network is introduced to achieve the purpose of feature transformation. Summary of the Invention

[0009] Purpose of the invention:

[0010] To overcome the limitations of current data-driven algorithms in the field of engine gas path fault diagnosis, which often hinder the construction of effective diagnostic models due to the difficulty in collecting fault data and the inability to build effective diagnostic models when there is little or no fault data, this paper applies domain adaptation methods from transfer learning to the field of turbofan engine gas path fault diagnosis. For cases where the target domain contains only normal category data, a partially aligned domain adversarial neural network is used to train a domain classifier using normal data from the source domain and target domain data (containing only normal categories), while a class classifier is trained using all source domain data, thereby constructing a high-performance fault diagnosis model.

[0011] Technical solution:

[0012] A method for diagnosing engine gas path faults based on a domain adversarial neural network comprises the following steps:

[0013] Step 1: Collect the state monitoring sensor data of the turbofan engine under different operating conditions in the degraded and non-degraded states and the corresponding fault states;

[0014] Step 2: Obtain the source domain and target domain for transfer learning based on sensor data. The source domain contains data of all categories, and the target domain contains only data of normal categories.

[0015] Step 3: Use the source domain data and the target domain data to train a domain adversarial neural network; the domain adversarial neural network includes a feature extractor, a category classifier, and a domain classifier; the objective function of the domain adversarial neural network is:

[0016]

[0017] Among them, the first term is the sample label prediction loss of the source domain data, the second term is the domain label prediction loss of the target domain data, and the third term is the domain label prediction loss of the normal category data in the source domain; L y (·,·) represents the prediction loss function of the sample label, L d (·,·) represents the prediction loss function of the domain label, λ is the coefficient of the gradient reversal layer, G f (·;θ f ) is the feature extractor, θ f Represents the parameters in the feature extractor, G y (·;θ y ) is the category classifier, θ y Represents the parameters in the category classifier, G d (·;θ d ) domain classifier, θ d Represents the parameters in the domain classifier, x i represents the training data, denote the source domain and target domain respectively, y iIndicates the fault category, d i Indicates the domain category.

[0018] Preferably, the turbofan engine status monitoring sensor data in step 1 includes high-pressure rotor speed, low-pressure rotor speed, fan outlet temperature, compressor outlet temperature, compressor outlet pressure, low-pressure turbine inlet temperature, low-pressure turbine outlet temperature, low-pressure turbine outlet pressure, mixing chamber inlet temperature, and main fuel flow.

[0019] Preferably, the aircraft engine failure state in step 1 includes fan failure, compressor failure, high-pressure turbine failure and low-pressure turbine failure.

[0020] Preferably, the degraded state in step 1 means that the state of the engine gas system will change due to external environmental interference and running time factors; the non-degraded state means that all components of the engine are in a standard state and can meet the design requirements; different working conditions mean that the engine inlet temperature and pressure will change with changes in the external environment, thereby affecting the working state of the engine.

[0021] Preferably, the feature extractor is a convolutional neural network structure, which adopts two layers of convolution, and the two layers of convolution respectively include a convolution layer, a batch normalization layer, a pooling layer and an activation layer.

[0022] Preferably, the category classifier and the domain classifier are both two-layer perceptron structures, including a fully connected layer, a BN layer, a Dropout layer, an activation layer and a fully connected layer.

[0023] Preferably, a gradient reversal layer is provided between the feature extractor and the domain classifier.

[0024] Preferably, the parameter θ of the domain classifier d By minimizing the domain classification loss, the parameters θ of the sample classifier are optimized. y The feature extractor parameters θ are optimized by minimizing the label prediction loss of the sample. f The optimization is performed by minimizing the sample label prediction loss and maximizing the domain classification loss. The formula is as follows:

[0025]

[0026] Preferably, step 2 includes: using a sliding window strategy with a window width of 10, converting ten one-dimensional data into one two-dimensional data, and a sliding step size of 1, that is, converting N one-dimensional data into (N-10+1) 10*10 two-dimensional data; then normalizing the sensor data and dividing the data under different working conditions in the same state into a group, which are used as the source domain and target domain of transfer learning respectively.

[0027] Preferably, both the category loss function and the domain loss function use the cross entropy loss function.

[0028] Beneficial effects:

[0029] This invention addresses the challenges of collecting fault data in turbofan engine gas path fault diagnosis, where the target domain contains only normal data. By applying domain adaptation methods from transfer learning to engine gas path fault diagnosis, this method constructs a high-performance target domain fault diagnosis model using a source domain containing data from all categories and a target domain containing only normal data. This method utilizes data from related but different domains to train the target domain model, addressing the problem of difficulty in acquiring fault data in the target domain, which results in poor diagnostic model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a structural diagram of the turbofan engine of the present invention.

[0031] Figure 2 It is a flow chart of the method of the present invention.

[0032] Figure 3 This is the test result graph. DETAILED DESCRIPTION

[0033] The present invention provides an engine gas path fault diagnosis method based on a domain adversarial neural network, comprising the following steps:

[0034] Step 1: Collect state monitoring sensor data for the turbofan engine under four different operating conditions, both in a degraded state and a non-degraded state, as well as the corresponding fault conditions. Turbofan engine state monitoring sensor data includes high-pressure rotor speed, low-pressure rotor speed, fan outlet temperature, compressor outlet temperature, compressor outlet pressure, low-pressure turbine inlet temperature, low-pressure turbine outlet temperature, low-pressure turbine outlet pressure, mixing chamber inlet temperature, and main fuel flow rate. Aircraft engine fault conditions include fan failure, compressor failure, high-pressure turbine failure, and low-pressure turbine failure. A degraded state refers to changes in the engine's airflow system due to external environmental interference (such as wind, sand, rain, and wear and corrosion) and operating time. For example, degradation of the compressor rotor seal reduces compressor boost capacity and efficiency, and fouling on compressor blades reduces flow rate. A non-degraded state indicates that all engine components are in standard condition and meet design requirements. Different operating conditions refer to changes in the engine inlet temperature and pressure due to external environmental factors, which in turn affect the engine's operating state. The main influencing factors are flight Mach number and altitude.

[0035] Step 2: Use a sliding window strategy with a window width of 10 to convert ten one-dimensional data into one two-dimensional data with a sliding step of 1, that is, N one-dimensional data are converted into (N-10+1) 10*10 two-dimensional data. Then normalize the sensor data and group the data under different working conditions in the same state into one group, which are used as the source domain and target domain of transfer learning respectively. The source domain contains data of all categories, and the target domain contains only data of normal categories. There are 24 groups of transfer tasks in total. The original data of each fault condition is 1500. After using a sliding window with a width of 10 and a sliding step of 1, 1491 10*10 two-dimensional data are obtained. The source domain contains data of all categories of normal conditions and four fault conditions, and the target domain contains only data of normal conditions. The sensor data normalization formula is as follows:

[0036]

[0037] Where x is the original sensor data, x norm is the normalized sensor data, μ and σ represent the mean and variance of the original sensor data distribution, respectively.

[0038] Step 3: Design the network structure. The feature extractor uses a two-layer convolutional structure. The first convolution layer includes a convolution layer, a batch normalization layer, a pooling layer, and an activation layer. The convolution layer in the first convolution layer has a convolution kernel size of 3*3 and 10 output channels. Next is the batch normalization layer (BN layer), which can accelerate the convergence of the network, prevent network overfitting, and prevent the gradient from vanishing. At the same time, because the network stores the variance and mean of the training data, it can reduce the difference between the source and target domains to a certain extent. After the BN layer is the pooling layer, the kernel size of this layer is 2*2. Next is the activation function. The activation function uses the ReLU function commonly used in deep learning. Its mathematical definition is shown below. This function has fast calculation speed and can prevent gradient vanishing. The convolution kernel size of the second convolution layer is 3*3 and the output channels are 40. The rest of the structure and parameters are the same as the first layer.

[0039] f(x)=max(0,x)

[0040] The category classifier primarily uses a two-layer perceptron. The first layer is a fully connected layer, connected to the output of the feature extractor. Therefore, its number of input nodes equals the number of feature extractor output points, which is 640, and its number of output nodes is set to 50. This is followed by a one-dimensional batch normalization layer, which also prevents network overfitting and improves generalization. This is followed by a dropout layer, which randomly inactivates some neurons during training to prevent the network from relying heavily on any one neuron, thereby preventing overfitting. This is followed by the activation function, again using the ReLU function, and finally a fully connected layer with 50 input nodes and 5 output nodes, equal to the number of sample categories. The domain classifier has a similar structure to the category classifier, with the final fully connected layer having 2 output nodes.

[0041] The network also has a gradient reversal layer. This layer does not have any parameters that need to be updated, but it does have a parameter λ that needs to be determined, which is determined as follows:

[0042]

[0043] Where P is the ratio of the current number of iterations to the total number of iterations.

[0044] Step 4: Put the source domain data and target domain data into the model for training. The objective function of the domain adversarial neural network is:

[0045]

[0046] Among them, L y (·,·) represents the prediction loss function of the sample label, L d (·,·) represents the prediction loss function of the domain label, λ is the coefficient of the gradient reversal layer, G f (·;θ f ) is the feature extractor, θ f Represents the parameters in the feature extractor, G y (·;θ y ) is the category classifier, θ y Represents the parameters in the category classifier, G d (·;θ d ) domain classifier, θ d Represents the parameters in the domain classifier, x i represents the training data, denote the source domain and target domain respectively, y i Indicates the fault category, d i Indicates the domain category.

[0047] Due to the particularity of engine data, the target domain only contains normal category data. Therefore, the source domain data and the target domain data cannot be directly used for domain adaptation. Instead, the source domain normal category data and the target domain normal data should be used for domain adaptation. Therefore, the objective function of the partially aligned domain adversarial neural network is (assuming that the sample label of the normal category data is 1):

[0048]

[0049] The first term is the sample label prediction loss of the source domain data, the second term is the domain label prediction loss of the target domain data, and the third term is the domain label prediction loss of the normal category data in the source domain.

[0050] Parameters θ of the domain classifier d By minimizing the domain classification loss, the parameters θ of the sample classifier are optimized. y The optimization is performed by minimizing the label prediction loss of the sample. The parameter θ of the feature extractor f Optimization is performed by minimizing the sample's label prediction loss and maximizing the domain classification loss. The formula is as follows:

[0051]

[0052] In step 5, according to the training results, the learning rate is adjusted to 1e-4, the training batch size is 100, the number of iterations is 50, and the category loss function and the domain loss function both use the cross entropy loss function.

[0053] Example

[0054] The data used in this case comes from a certain type of turbofan engine, whose structure is as follows Figure 1 As shown. There are 10 characteristic parameters: high-pressure rotor speed, low-pressure rotor speed, fan outlet temperature, compressor outlet temperature, compressor outlet pressure, low-pressure turbine inlet temperature, low-pressure turbine outlet temperature, low-pressure turbine outlet pressure, mixing chamber inlet temperature, and main fuel flow rate. The engine is divided into a degraded state and a non-degraded state. In the normal state, the compressor and turbine of the engine are at the design performance, that is, the efficiency and flow rate are both design values; in the degraded state, the compressor efficiency decreases and the flow rate becomes smaller, and the turbine efficiency decreases and the flow rate becomes larger. The specific changes are shown in Table 1. Engine air path failures are divided into four categories: fan failure, compressor failure, high-pressure turbine failure and low-pressure turbine failure. The corresponding health parameter changes under each type of failure are shown in Table 2. In addition, taking into account the impact of different external conditions on engine performance and the generalization performance of the algorithm, different external conditions are obtained by changing the altitude and forward flight speed, as shown in Table 3. The specific network structure of the present invention is shown in Table 4, and the specific method flow chart is shown in Table 4. Figure 2 shown.

[0055] To illustrate the effectiveness of the algorithm, we use accuracy as an indicator to measure the performance of the algorithm, that is, the ratio of the number of correctly predicted samples to the number of predicted samples. For comparison, we set up four benchmark algorithms for comparison:

[0056] 1) BP1: Experiments were conducted using a single-hidden-layer BP algorithm. The hidden layer had 10 nodes, the input was a vector with 10 feature dimensions, the output was five data categories, and the activation function was ReLU. The training data consisted of all data in the source domain, and the test data consisted of all data in the target domain.

[0057] 2) BP2: The network is a single-hidden-layer BP neural network. The hidden layer has 10 nodes, the input node is a vector of 10 feature dimensions, the output is five data categories, and the activation function is ReLU. The training data consists of all data in the source domain and normal data in the target domain, and the test data consists of all data in the target domain.

[0058] 3) CNN: The network structure is the same as the feature extractor plus sample classifier in Table 1. The training data is all the data in the source domain and the normal data in the target domain, and the test data is all the data in the target domain.

[0059] 4) DANN: The network structure is exactly the same as that shown in Table 1, but the data calculation in the network is to align the data in the target domain with all the data in the source domain.

[0060] The specific migration tasks are shown in Table 5, and the comparison results are shown in Table 6 and Table 7, where PADANN is the method provided by the present invention. The comparison of the results shows that compared with several benchmark algorithms, the method provided by the present invention achieves excellent gas path fault diagnosis results when there is only normal category data in the target domain and more source domain data. The test result histogram is shown in Figure 5. Figure 3 shown.

[0061] Table 1 Changes in engine health parameters under degradation conditions

[0062]

[0063] Table 2 Engine health parameters under various fault conditions

[0064]

[0065] Table 3 External conditions

[0066]

[0067] Table 4 Network structure

[0068]

[0069] Table 5 Migration task details

[0070]

[0071] Table 6 Mean and standard deviation of diagnostic accuracy of group A experiments

[0072]

[0073] Table 7 Mean and standard deviation of diagnostic accuracy of group B experiments

[0074]

Claims

1. A method for diagnosing engine gas path faults based on a domain adversarial neural network, characterized in that: The steps include: Step 1: Collect the state monitoring sensor data of the turbofan engine under different operating conditions in the degraded and non-degraded states and the corresponding fault states; Step 2: Obtain the source domain and target domain for transfer learning based on sensor data. The source domain contains data of all categories, and the target domain contains only data of normal categories. Step 3: Use the source domain data and target domain data to train the domain adversarial neural network; The domain adversarial neural network includes a feature extractor, a category classifier, and a domain classifier; a gradient reversal layer is set between the feature extractor and the domain classifier; the objective function of the domain adversarial neural network is: Among them, the first term is the sample label prediction loss of the source domain data, the second term is the domain label prediction loss of the target domain data, and the third term is the domain label prediction loss of the normal category data in the source domain; L y (·,·) represents the prediction loss function of the sample label, L d (·,·) represents the prediction loss function of the domain label, λ is the coefficient of the gradient reversal layer, G f (·;θ f ) is the feature extractor, θ f Represents the parameters in the feature extractor, G y (·;θ y ) is the category classifier, θ y Represents the parameters in the category classifier, G d (·;θ d ) domain classifier, θ d Represents the parameters in the domain classifier, x i represents the training data, denote the source domain and target domain respectively, y i Indicates the fault category, d i Indicates the domain category.

2. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The turbofan engine status monitoring sensor data in step 1 includes high-pressure rotor speed, low-pressure rotor speed, fan outlet temperature, compressor outlet temperature, compressor outlet pressure, low-pressure turbine inlet temperature, low-pressure turbine outlet temperature, low-pressure turbine outlet pressure, mixing chamber inlet temperature, and main fuel flow.

3. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The aircraft engine fault states in step 1 include fan fault, compressor fault, high-pressure turbine fault and low-pressure turbine fault.

4. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The degraded state in step 1 means that the state of the engine's air path system will change due to external environmental interference and operating time factors; the non-degraded state means that all engine components are in standard condition and can meet design requirements; different operating conditions mean that the engine inlet temperature and pressure will change with changes in the external environment, thereby affecting the engine's operating state.

5. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The feature extractor is a convolutional neural network structure, which adopts two layers of convolution, and the two layers of convolution include a convolution layer, a batch normalization layer, a pooling layer and an activation layer respectively.

6. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The category classifier and the domain classifier both have a two-layer perceptron structure, including a fully connected layer, a BN layer, a Dropout layer, an activation layer, and a fully connected layer.

7. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The parameters θ of the domain classifier d By minimizing the domain classification loss, the parameters θ of the sample classifier are optimized. y Optimization is performed by minimizing the label prediction loss of the samples.

8. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: The parameters θ of the feature extractor f The optimization is performed by minimizing the sample’s label prediction loss and maximizing the domain classification loss.

9. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1 is characterized in that: Step 2 includes: using a sliding window strategy to convert one-dimensional data into two-dimensional data; then normalizing the sensor data and grouping data under different working conditions under the same state into a group, which serves as the source domain and target domain of transfer learning respectively.

10. The engine gas path fault diagnosis method based on domain adversarial neural network according to claim 1, characterized in that: Both the category loss function and the domain loss function use the cross entropy loss function.

Images