An aero-engine abnormality diagnosis method, device, medium and product

Abstract

The application discloses an aero-engine abnormality diagnosis method, device, medium and product, relates to the technical field of aero-engine health monitoring, and determines reconstructed data based on an adversarial multi-head attention network and current observation data of an aero-generator; inputs the reconstructed data into a space-time feature enhancement module to determine a space-time feature vector; takes the space-time feature vector as a current state, and determines a diagnosis result corresponding to the current state based on a reinforcement learning diagnosis model; and determines an abnormality early warning result of the aero-engine based on the diagnosis result corresponding to the current state and a preset fault type set. The application can improve the accuracy of aero-engine abnormality diagnosis.

Description

Technical Field

[0001] This application relates to the field of aircraft engine health monitoring technology, and in particular to an aircraft engine anomaly diagnosis method, equipment, medium and product. Background Technology

[0002] As the core power unit of an aircraft, the real-time monitoring and anomaly diagnosis of the aero-engine's operating status are crucial for flight safety. Traditional diagnostic methods mainly rely on the complete data sequence collected in real time by engine sensors, using threshold judgment, spectrum analysis, or air path analysis based on physical models to locate faults. However, the engine's operating environment is extremely harsh, and sensors are susceptible to factors such as vibration, electromagnetic interference, and high temperatures, leading to frequent issues of missing partial sensor data during data acquisition. For example, a compressor outlet temperature sensor may lose data from multiple consecutive sampling points due to instantaneous overload. Such data loss severely damages signal integrity, rendering traditional diagnostic models based on complete data ineffective.

[0003] Existing solutions to the data missing problem mainly fall into two categories: one is data imputation-based preprocessing methods, such as linear interpolation and K-nearest neighbor imputation. While these methods can easily fill in missing values, they ignore the complex temporal correlations and spatial couplings between aero-engine parameters, leading to biases in the reconstructed data. The other category is intelligent algorithms that directly process missing data, such as data reconstruction models combining generative adversarial networks (GANs) with attention mechanisms, and degradation state prediction methods based on transfer learning. However, GANs are prone to pattern collapse in scenarios with small sample data, and the weight allocation of attention mechanisms lacks dynamic adjustment capabilities; while transfer learning requires a large amount of labeled data, making it difficult to adapt to the actual needs of scarce aero-engine fault samples. Therefore, there is an urgent need to propose a method for aero-engine anomaly diagnosis when some sensor data is missing. Summary of the Invention

[0004] The purpose of this application is to provide a method, device, medium, and product for diagnosing aero-engine anomalies, which can improve the accuracy of aero-engine anomaly diagnosis.

[0005] To achieve the above objectives, this application provides the following solution.

[0006] Firstly, this application provides a method for diagnosing anomalous conditions in an aircraft engine, including the following:

[0007] Based on current observational data from the adversarial multi-head attention network and the aircraft generator, reconstructed data is determined. This observational data comprises multiple sensor data containing missing values, determined from different engine sensors.

[0008] The reconstructed data is input into the spatiotemporal feature enhancement module to determine the spatiotemporal feature vector; the spatiotemporal feature enhancement module is determined based on dilated convolution and bidirectional long short-term memory network.

[0009] The spatiotemporal feature vector is used as the current state, and the diagnostic result corresponding to the current state is determined based on the reinforcement learning diagnostic model. The diagnostic result corresponding to the current state is the fault type corresponding to the current state of the aero-engine. The reinforcement learning diagnostic model is determined based on the gradient of a dual-delay deep deterministic policy.

[0010] Based on the diagnostic results corresponding to the current state and the preset set of fault types, the abnormal early warning results of the aero-engine are determined.

[0011] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for diagnosing anomalous conditions in an aero-engine.

[0012] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned aero-engine anomaly diagnosis method.

[0013] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for diagnosing anomalous conditions in an aero-engine.

[0014] Based on the specific embodiments provided in this application, the following technical effects are disclosed.

[0015] This application first determines reconstructed data based on adversarial multi-head attention networks and current observation data of aero-engines, achieving high-precision reconstruction of sensor data containing missing values. Then, the reconstructed data is input into a spatiotemporal feature enhancement module to determine the spatiotemporal feature vector, which is determined based on dilated convolution and a bidirectional long short-term memory network, achieving time-frequency domain representation of anomalous signals. Further, the spatiotemporal feature vector is used as the current state, and a reinforcement learning diagnostic model is used to determine the corresponding diagnostic result. The reinforcement learning diagnostic model is based on gradient determination using a dual-delay deep deterministic strategy, achieving adaptive optimization of the diagnostic strategy. The reinforcement learning diagnostic model significantly improves the robustness of diagnosis in the case of missing sensor data. Finally, based on the diagnostic result and a preset set of fault types, an aero-engine anomaly warning result is determined. This application effectively improves the accuracy of aero-engine anomaly diagnosis and provides a reliable technical solution for real-time health monitoring of aero-engines. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an aero-engine anomaly diagnosis method provided in the embodiments of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this application provides a method for diagnosing anomalous conditions in an aircraft engine, including the following:

[0021] Step 11: Determine the reconstructed data based on the current observation data of the adversarial multi-head attention network and the aero-engine; the observation data is multiple sensor data containing missing values ​​determined based on different engine sensors.

[0022] In some embodiments, step 11 specifically includes: constructing a generative adversarial multi-head attention network based on a multi-head attention network, a generator, and a discriminator; based on the generator, extracting local features of the observed data using convolutional layers, and determining the spatial and temporal dependencies between different sensor data through the multi-head attention network, and outputting initial reconstructed data; determining the discriminant output for the observed data and the discriminant output for the initial reconstructed data according to the discriminator; determining the loss function of the discriminator based on the discriminant output for the observed data and the discriminant output for the initial reconstructed data; updating the initial reconstructed data according to the loss function of the discriminator, and determining the final reconstructed data.

[0023] Step 12: Input the reconstructed data into the spatiotemporal feature enhancement module to determine the spatiotemporal feature vector; the spatiotemporal feature enhancement module is determined based on dilated convolution and bidirectional long short-term memory network.

[0024] In some embodiments, step 12 specifically includes: inputting the reconstructed data into a dilated convolution, capturing the features of the reconstructed data based on dilated convolutions with different dilation rates, and determining multi-scale temporal features; inputting the multi-scale temporal features into a bidirectional long short-term memory network, and outputting a spatiotemporal feature vector.

[0025] Step 13: Take the spatiotemporal feature vector as the current state, and determine the diagnosis result corresponding to the current state based on the reinforcement learning diagnostic model; wherein, the diagnosis result corresponding to the current state is the fault type corresponding to the current state of the aero-engine; the reinforcement learning diagnostic model is determined based on the gradient of the dual-delay deep deterministic policy.

[0026] In some embodiments, step 13 specifically includes steps 21-25.

[0027] Step 21: The reinforcement learning diagnostic model includes an Actor policy network and a dual-Q network; the dual-Q network consists of two Critic networks.

[0028] Step 22: Input the current state into the Actor policy network to determine the initial diagnostic result corresponding to the current state.

[0029] In some embodiments, step 22 specifically includes: inputting the current state into the Actor policy network to determine the action probability distribution corresponding to the initial action; and selecting the initial action corresponding to the maximum value in the action probability distribution as the initial diagnostic result.

[0030] Step 23: Take the initial diagnosis result corresponding to the current state as the current action, and take the current state and the current action as the current state-action pair.

[0031] Step 24: Evaluate the current state action pair based on two Critic networks to determine the target Q value.

[0032] In some embodiments, step 24 specifically includes: evaluating the current state-action pair according to a Critic network to determine a first target Q value; evaluating the current state-action pair according to another Critic network to determine a second target Q value; selecting the minimum of the first target Q value and the second target Q value, and determining a target Q value based on the reward corresponding to the current action.

[0033] Step 25: With the goal of maximizing the target Q value, update the Actor policy network and determine the updated current action as the diagnosis result corresponding to the current state.

[0034] In some embodiments, the loss function of the discriminator is as follows.

[0035] .

[0036] in, This is real data obtained from different engine sensors; To reconstruct the data; For the discriminator to reconstruct data The discriminant output; The discriminator outputs its judgment on the real data X. This represents the true data distribution. For the distribution of observed data; The expected value of the real data sampled from the real data distribution obtained from different engine sensors; This represents the expected value of the observation data sampled from the observation data distribution. Let be the loss function of the discriminator.

[0037] Step 14: Based on the diagnostic results corresponding to the current state and the preset set of fault types, determine the abnormal warning result of the aero-engine.

[0038] In practical applications, the specific process of the aero-engine anomaly diagnosis method in the case of missing sensor data includes the following steps.

[0039] Step 100: Aero-engine sensor data exhibits strong spatial coupling and temporal dependence. Traditional interpolation methods cannot preserve this complex correlation, leading to biases in subsequent diagnostic models. Therefore, a Generative Adversarial Network (GAN-MAN) is constructed to achieve high-precision recovery of missing data. This module consists of a generator G and a discriminator D. The generator employs parallel convolutional layers and a multi-head attention mechanism to capture the spatial correlation and temporal dependence of sensor data. Specifically, after standardization, the observed data X is concatenated with a mask matrix M (obtained by marking the missing positions of each element in the observed data; missing positions are marked as 0, and non-missing positions as 1). Local features are extracted through convolutional layers, and then the dynamic correlation weights between different sensors are calculated through the multi-head attention layer. The generator's output is the reconstructed data. Its expression is as follows.

[0040] 。

[0041] in, For the observed non-missing data, M It is a mask matrix.

[0042] Step 101, Discriminator DThen, through adversarial training, it is determined whether the input data is real data or reconstructed data generated by the generator, and the loss function is used. as follows.

[0043] .

[0044] in, This is real data obtained from different engine sensors; To reconstruct the data; For the discriminator to reconstruct data The discriminant output; The discriminator outputs its judgment on the real data X. This represents the true data distribution. For the distribution of observed data; The expected value of the real data sampled from the real data distribution obtained from different engine sensors; This represents the expected value of the observation data sampled from the observation data distribution. Let be the loss function of the discriminator.

[0045] This design transforms the reconstruction of missing data (missing values) in the observation data into a conditional generation problem. It uses adversarial game theory to force the generator to learn the distribution characteristics of real data. At the same time, it explicitly models the dynamic correlation between sensors through a multi-head attention mechanism, thus solving the problem of insufficient reconstruction accuracy of traditional methods under complex working conditions.

[0046] Step 200: Although the reconstructed data recovers the missing values, the time-frequency domain features of the original signal still need further enhancement. The reconstructed data is input into the spatiotemporal feature enhancement module, which employs a cascaded structure of dilated convolution and a bidirectional long short-term memory (BiLSTM) network to achieve joint extraction of multi-scale temporal features and bidirectional temporal dependencies. Dilated convolution expands the receptive field and captures multi-scale temporal features without increasing the number of parameters through dilated convolutions with different dilation rates. l The output of the layer dilated convolution is the multi-scale temporal feature extracted by the current layer dilated convolution. The formula is as follows.

[0047] .

[0048] in, This is a one-dimensional convolution operation. For the first The input to the layer dilated convolution is k, where k is the kernel size and s is the stride. For the first The dilation rate of each layer of dilated convolution. First layer dilated convolution. The input comes directly from the module's initial input, i.e., the reconstructed data. For the first layer( The dilated convolution of ) with input It is the first The output of layer dilated convolution.

[0049] Step 201: Input the data after dilated convolution into BiLSTM, forward hidden state. and backward hidden state From the current input It is calculated from the hidden state at the previous time step.

[0050] .

[0051] .

[0052] in, Indicates a forward LSTM unit. Indicates a backward LSTM unit. For the forward LSTM at time step The hidden state, For backward LSTM at time step The hidden states are defined. The output feature vector F is obtained by concatenating the bidirectional hidden states and then reducing its dimensionality through a fully connected layer. It contains information on time-domain trends and frequency-domain fluctuations, and its dimension is... d × T , where d is the feature dimension and T is the time step.

[0053] Step 300: The anomaly diagnosis of aero-engines is essentially a sequential decision-making problem, requiring a balance between diagnostic accuracy, false alarm rate, and false negative rate under the uncertainty of features caused by missing data. Traditional supervised learning relies on fixed thresholds, making it difficult to adapt to changes in operating conditions. The improved TD3 algorithm, however, suppresses overestimation through a double Q-network and reduces variance through smoothing the objective policy, making it suitable for decision optimization in high-risk scenarios. A reinforcement learning diagnostic model based on Twin Delayed Deep Deterministic Policy Gradient (TD3) is designed. The state space is defined as the spatiotemporal feature vector F, and the action space is a discrete set of diagnostic decisions. ,in Let R represent the i-th exception type. The reward function R is designed as follows.

[0054] .

[0055] in, , , These are the weighting coefficients. , False alarm rate This represents the false negative rate. Specifically... The proportion of samples where the model correctly diagnoses an anomaly type, i.e., the percentage of samples where the predicted anomaly type matches the actual anomaly type. The false alarm rate is the proportion of normal states that the model misclassifies as abnormal states. The false negative rate is the proportion of states where the model fails to detect actual anomalies.

[0056] Step 301: Store the current state s, the current action a, the reward r corresponding to the current action, and the next state s' in the experience replay pool, using a dual... Network computing objectives value.

[0057] .

[0058] in, It is Gaussian noise; This is the target policy network, i.e., the updated Actor policy network; Represents the value function of two target actions and Choose the smaller value. This indicates the next state in the target network. Take action below The value of the action is estimated. For the target network; For serial numbers. Specifically, based on the target. Value update current double network( and ) to obtain the target double Network (target network, including) and ).

[0059] To address the Q-value overestimation problem in traditional deep reinforcement learning algorithms, the TD3 algorithm employs a dual-Q network, where the current Critic network comprises two independent Critic networks, denoted as... and The target Critic network (target network) refers to the network with two target Critic networks for each current Critic network, denoted as . and .

[0060] Step 302: The purpose of the Actor policy network is to output the optimal diagnostic strategy to achieve accurate judgment of abnormal conditions of the aero-engine. The Actor policy network updates the strategy by maximizing the Q-value as follows.

[0061] .

[0062] in, The objective function of the Actor policy network Regarding parameters The gradient. Indicates the state distribution Next, sample the current state s and calculate the corresponding expectation. Action value function The gradient, and the current action Fixed as Actor policy network The output action. Represents the Actor policy network Regarding parameters The gradient; These are the trainable parameters of a Critic network. The target action value function. and It is determined based on two target Critic networks; action value function It is determined based on a Critic network.

[0063] It is worth noting that the TD3 algorithm uses a dual... network( and The core purpose is to solve The problem of overestimation is specifically reflected in the target value calculation stage (using...). (Suppressing overestimation), but in the Actor policy network update phase, usually only one Q-function is needed to guide the policy gradient; in practical applications, alternatives can also be chosen. The key is to maintain consistency.

[0064] This decision-making mechanism models the diagnostic process as a Markov decision process and uses reinforcement learning to automatically optimize strategies from historical diagnostic experience. It solves the problems of traditional methods that rely on manual parameter tuning and have weak generalization ability, and is especially suitable for robust diagnosis in small-sample fault scenarios.

[0065] Step 303: The Actor policy network outputs the action probability distribution based on the current state s. Choose the action with the highest probability. As a diagnostic result. If Belongs to the preset set of exception types ,in, The number of aircraft engine failure types; for If a preset type of aircraft engine fault is identified (such as compressor failure, turbine overheating, etc.), an abnormality is determined, triggering the corresponding alarm mechanism, such as sending an alarm message to the ground control center or issuing an audible and visual alarm in the aircraft cockpit; if If it does not belong to the abnormal type set, then the engine is determined to be operating normally.

[0066] At the same time, during the diagnostic process, the current state s and the current action are recorded each time. The reward r corresponding to the current action (based on the reward function and the next state) Data is stored in the experience replay pool. The reinforcement learning diagnostic model is periodically updated and trained using randomly sampled data from the experience replay pool to adapt to changes in the operating status of the aero-engine and newly emerging anomaly patterns.

[0067] By integrating data reconstruction, feature enhancement, and reinforcement learning decision-making, robust compensation for missing data and accurate identification of abnormal patterns can be achieved.

[0068] Example 2

[0069] This application can be achieved through the following technical solutions, the specific solutions are as follows.

[0070] 1. Multimodal data reconstruction module.

[0071] Aero-engine sensor data exhibits strong spatial coupling and temporal dependence, making it difficult for traditional interpolation methods to preserve this complex correlation, leading to biases in subsequent diagnostic models. To address this, a GAN-MAN module was designed.

[0072] Generator (G): Receives standardized observation data Using a mask matrix M (marking missing data locations), local features from individual sensors are extracted through parallel convolutional layers. Then, a multi-head attention mechanism is used to calculate dynamic correlation weights across sensors, achieving context-aware reconstruction of missing data and outputting complete data that closely approximates the true distribution. .

[0073] Discriminator (D): Through adversarial training, it distinguishes between real and generated data, forcing the generator to learn the distribution characteristics of real data. This module transforms missing data reconstruction into a conditional generation problem, solving the problem of insufficient reconstruction accuracy of traditional methods under complex conditions.

[0074] 2. Spatiotemporal feature enhancement module.

[0075] Although the reconstructed data recovered the missing values, the anomalous features implicit in the original signal still needed to be enhanced. A cascaded structure of dilated convolution and bidirectional long short-term memory network (BiLSTM) was adopted.

[0076] Dilated convolutional layers: By expanding the receptive field through dilated convolutions with different dilation rates, multi-scale temporal features can be captured without increasing the number of parameters. Hierarchical information is extracted layer by layer from short-term fluctuations (small dilation rate) to long-term trends (large dilation rate).

[0077] BiLSTM layer: Performs bidirectional temporal modeling on the feature sequence output by dilated convolution. Forward LSTM captures the influence of past information on the current state, and backward LSTM mines the dependence of future information on the current state. Finally, the bidirectional hidden states are concatenated and the dimensionality is reduced to obtain a spatiotemporal feature vector F containing temporal trends and frequency domain fluctuations, providing a highly robust input for subsequent decision-making.

[0078] 3. A reinforcement learning diagnostic module based on an improved TD3.

[0079] The diagnosis of aero-engine anomalies is essentially a sequential decision-making problem, requiring a balance between diagnostic accuracy, false alarm rate, and false negative rate under the uncertainty of features caused by missing data. Traditional supervised learning relies on fixed thresholds, which are difficult to adapt to changes in operating conditions. The improved TD3 algorithm achieves intelligent decision-making through the following mechanism.

[0080] State and Action Definition: Using the spatiotemporal feature vector F as the state space and the discrete anomaly type set as the action space, a multi-index reward function is designed, which includes diagnostic accuracy, false alarm rate, and false negative rate. The diagnostic needs of different operating stages are dynamically balanced through weight coefficients.

[0081] pair Network and target strategy: Employing a dual approach Network computing objectives The value is minimized to suppress overestimation; Gaussian noise is added to the output of the target policy network to enhance the exploratory nature of actions and avoid getting trapped in local optima.

[0082] Policy Update: The Actor policy network maximizes... The value update strategy learns the optimal decision from historical diagnostic experience and models the diagnostic process as a Markov decision process, which solves the problems of traditional methods that rely on manual parameter tuning and have weak generalization ability. It is especially suitable for small sample failure scenarios.

[0083] 4. Online diagnosis and model updates.

[0084] After real-time acquisition of sensor data, the spatiotemporal feature vector is obtained through data reconstruction and feature enhancement. The vector is then input into the reinforcement learning model to output diagnostic results. If the abnormality belongs to a preset type, a multi-level alarm is triggered. At the same time, the diagnostic experience is stored in the playback pool, and the model is updated regularly to adapt to engine performance degradation or new fault modes.

[0085] Compared with the prior art, this application has the following beneficial effects.

[0086] First, this application effectively captures the dynamic correlation between sensors by generating an adversarial multi-head attention network, solving the problem of traditional interpolation methods destroying data coupling relationships and making the reconstructed data closer to the real distribution. By combining dilated convolution and BiLSTM to extract multi-scale spatiotemporal features, it strengthens the time-frequency domain representation of abnormal signals, provides high-discrimination input for diagnostic models, and significantly improves feature utilization and diagnostic accuracy in scenarios with missing data.

[0087] Second, this application, based on the improved TD3 reinforcement learning model, balances diagnostic accuracy, false alarm rate, and false negative rate through a multi-index reward function, utilizing a dual-index reward function. The network suppresses overestimation and enhances the exploratory nature of the target policy by reducing noise, thus achieving adaptive optimization of the diagnostic strategy. This mechanism breaks through the limitations of traditional fixed thresholds, especially in small-sample fault scenarios. It can automatically learn the optimal decision through historical experience, reduce the cost of manual parameter tuning, and significantly improve the diagnostic reliability and generalization ability under extreme conditions.

[0088] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the methods described above.

[0089] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the methods described above.

[0090] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the methods described above.

[0091] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0092] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRdM), magnetic random access memory (MRdM), ferroelectric random access memory (FRdM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RdM) or external cache memory, etc. By way of illustration and not limitation, RdM can take many forms, such as static random access memory (SRdM) or dynamic random access memory (DRdM).

[0093] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for diagnosing anomalous conditions in an aircraft engine, characterized in that, include: Based on current observational data of adversarial multi-head attention networks and aero-engines, reconstructed data is determined; The observation data is multiple sensor data containing missing values ​​determined based on different engine sensors; the reconstructed data is input into the spatiotemporal feature enhancement module to determine the spatiotemporal feature vector; specifically, this includes: inputting the reconstructed data into dilated convolution, capturing the features of the reconstructed data based on dilated convolution with different dilation rates to determine multi-scale temporal features; inputting the multi-scale temporal features into a bidirectional long short-term memory network to output the spatiotemporal feature vector; the spatiotemporal feature enhancement module includes dilated convolution and a bidirectional long short-term memory network; The spatiotemporal feature vector is used as the current state, and the diagnostic result corresponding to the current state is determined based on a reinforcement learning diagnostic model. The diagnostic result corresponding to the current state is the fault type corresponding to the current state of the aero-engine. The reinforcement learning diagnostic model is determined based on a dual-delay deep deterministic policy gradient. The state space of the reinforcement learning diagnostic model is the spatiotemporal feature vector F, and the action space is a discrete set of diagnostic decisions. ,in Let R represent the i-th anomaly type; the reward function R of the reinforcement learning diagnostic model is designed as follows: ； in, , , These are the weighting coefficients. , False alarm rate The false alarm rate is the percentage of normal states that the model correctly diagnoses abnormality types; the false alarm rate is the percentage of normal states that the model misclassifies as abnormal states; and the false alarm rate is the percentage of true abnormal states that the model fails to detect. Based on the diagnostic results corresponding to the current state and the preset set of fault types, the abnormal early warning results of the aero-engine are determined.

2. The method for diagnosing aero-engine anomalies according to claim 1, characterized in that, Based on current observational data from adversarial multi-head attention networks and aero-engines, reconstructed data is determined, specifically including: A generative adversarial multi-head attention network is constructed based on a multi-head attention network, a generator, and a discriminator. Based on the generator, local features of the observation data are extracted using convolutional layers, and the spatial and temporal dependencies between different sensor data are determined through a multi-head attention network before the initial reconstructed data is output. Based on the discriminator, determine the discriminant output for the observed data and the discriminant output for the initial reconstructed data; The loss function of the discriminator is determined based on the discriminant output of the observed data and the discriminant output of the initial reconstructed data. The initial reconstructed data is updated based on the discriminator's loss function to determine the final reconstructed data.

3. The method for diagnosing aero-engine anomalies according to claim 1, characterized in that, Using the spatiotemporal feature vector as the current state, and based on a reinforcement learning diagnostic model, the diagnostic result corresponding to the current state is determined, specifically including: The reinforcement learning diagnostic model includes an Actor policy network and a dual-Q network; the dual-Q network consists of two Critic networks. The current state is input into the Actor policy network to determine the initial diagnostic result corresponding to the current state. The initial diagnosis result corresponding to the current state is taken as the current action, and the current state and the current action are taken as the current state-action pair; The current state-action pair is evaluated using two Critic networks to determine the target Q value; With the goal of maximizing the target Q value, the Actor policy network is updated, and the updated current action is determined as the diagnosis result corresponding to the current state.

4. The method for diagnosing aero-engine anomalies according to claim 3, characterized in that, The current state is input into the Actor policy network to determine the initial diagnostic result corresponding to the current state, specifically including: Input the current state into the Actor policy network to determine the probability distribution of the action corresponding to the initial action; The initial action corresponding to the maximum value in the action probability distribution is selected as the initial diagnostic result.

5. The method for diagnosing aero-engine anomalies according to claim 3, characterized in that, The current state-action pair is evaluated using two Critic networks to determine the target Q-value, specifically including: The first objective Q-value is determined by evaluating the current state-action pair using a Critic network. The second objective Q-value is determined by evaluating the current state action pair using another Critic network; Choose the minimum of the first target Q value and the second target Q value, and determine the target Q value based on the reward corresponding to the current action.

6. The method for diagnosing aero-engine anomalies according to claim 2, characterized in that, The loss function of the discriminator is: ； in, This is real data obtained from different engine sensors; To reconstruct the data; For the discriminator to reconstruct data The discriminant output; The discriminator outputs its judgment on the real data X. This represents the true data distribution. For the distribution of observed data; The expected value of the real data sampled from the real data distribution obtained from different engine sensors; This represents the expected value of the observation data sampled from the observation data distribution. Let be the loss function of the discriminator.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the aero-engine anomaly diagnosis method according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the aero-engine anomaly diagnosis method according to any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the aero-engine anomaly diagnosis method according to any one of claims 1-6.

Images