Method, device and equipment for predicting life of high-pressure turbine blade in aero-engine

Abstract

This application relates to the field of risk detection technology, and discloses a method, apparatus, and equipment for predicting the lifespan of high-pressure turbine blades in aero-engines. The method includes: statistically analyzing multidimensional factor parameters related to the aero-engine and decomposing these parameters into multiple feature parameter groups; acquiring a data sample set of the feature parameter groups and selecting a parameter model suitable for each group; training and validating the parameter model using the data sample set, and adding parameter models whose validation results meet preset conditions to a model cache pool; selecting representative models from the model cache pool and using these representative models to perform multidimensional predictions on actual data from a target engine; and generating a risk prediction level for the target engine. The technical solution provided by this application can accurately predict the risk level of an engine, thereby determining the lifespan of the high-pressure turbine blades.

Description

Technical Field

[0001] This application relates to the field of risk detection technology, and in particular to a method, apparatus and equipment for predicting the life of high-pressure turbine blades in aero engines. Background Technology

[0002] A civil aircraft engine consists of a fan, a low-pressure compressor, a high-pressure compressor, a combustion chamber, a high-pressure turbine, and a low-pressure turbine. After the gas mixes and expands with fuel in the combustion chamber, it first passes through the high-pressure turbine (HPTB), which drives the preceding high-pressure compressor. Then it passes through the low-pressure turbine, which drives the low-pressure compressor and the fan. The HPTB is arguably the engine's most critical and core power source; if it fails, the high-pressure compressor ceases to function, the engine loses all power, and the aircraft shuts down in flight.

[0003] Conventional civil aircraft engines are inspected using borescope inspection, which involves inserting an industrial endoscope into the engine and displaying the image via a camera for inspection by specialized engineers. However, the critical part of the HPTB component is the bottom connecting tenon, which cannot be directly seen and cannot be inspected using borescope inspection.

[0004] Therefore, current technology cannot detect the condition of the high-pressure turbine blade root while the engine is in a wing-mounted state. The lifespan of components can only be controlled by relying on the service life (number of flight cycles) provided by the component manufacturer. Clearly, this control method cannot accurately determine the operating status of the high-pressure turbine blades, potentially leading to the risk of in-flight engine shutdown. Summary of the Invention

[0005] This application provides a method, apparatus, and equipment for predicting the lifespan of high-pressure turbine blades in an aero-engine, which can accurately predict the risk level of the engine without disassembling the engine, and thus determine the lifespan of the high-pressure turbine blades.

[0006] In one aspect, this application provides a method for predicting the lifespan of high-pressure turbine blades in an aero-engine. The method includes: statistically analyzing multidimensional factor parameters related to the aero-engine, and decomposing the multidimensional factor parameters into multiple feature parameter groups. The multidimensional factor parameters are used to characterize engine performance parameters and / or flight status parameters. Each feature parameter group includes one or more factor parameters. For any feature parameter group, a data sample set of the feature parameter group is obtained, and a parameter model adapted to the feature parameter group is selected based on the number of factor parameters contained in the feature parameter group. Wherein, if the number of factor parameters contained in the feature parameter group is more than one, the selected parameter model... The method includes at least a parallel network branch for performing weight splitting; the parameter model is trained and validated using the data sample set, and parameter models whose validation results meet preset conditions are added to the model cache pool; representative models are selected from the model cache pool for each feature parameter group, and the selected representative models are used to perform multi-dimensional predictions on the actual data of the target engine to obtain multiple single-dimensional prediction results; based on the weight coefficients assigned to each feature parameter group, the multiple single-dimensional prediction results are weighted and summed to generate a risk prediction level for the target engine, wherein the risk prediction level is used to characterize the service life of the high-pressure turbine blades in the target engine.

[0007] The technical solution provided in this embodiment decomposes multidimensional factor parameters related to aero-engines into multiple characteristic parameter groups. By training models on each characteristic parameter group, individual parameter models for each group can be obtained. During training, different parameter models with varying structures can be selected based on the number of factor parameters contained in each characteristic parameter group, making the feature processing more accurate. Subsequently, by selecting representative models for each characteristic parameter group, different single-dimensional prediction results can be obtained using these representative models. By weighted summing of these single-dimensional prediction results, an accurate risk prediction level can be obtained, which can be used to characterize the service life of high-pressure turbine blades. The technical solution provided in this embodiment enables accurate prediction of engine risk levels without engine disassembly, thereby determining the service life of high-pressure turbine blades and effectively avoiding in-flight engine shutdowns.

[0008] In one implementation, training and validating the parameter model using the data sample set, and adding the parameter model whose validation results meet preset conditions to the model cache pool includes: extracting a first proportion of data from the data sample set as a global validation set, the global validation set including data samples representing healthy states and data samples representing unhealthy states; training the parameter model in specified batches, each batch of training including a specified number of iterative training sessions, and each iterative training session including a specified number of training rounds; wherein, for any batch of training, a second proportion of data from the data sample set other than the global validation set is selected as a local validation dataset, and in any iterative training session, a third proportion of data is selected as a local validation dataset. The test dataset is used, and a fourth proportion of data is selected as the training dataset. During each training round, the parameter model is trained using the training dataset, and after training, the trained parameter model is tested using the test dataset to determine the gradient descent direction. After completing all training and testing rounds in a set of iterative training, a new test dataset and training dataset are selected, and the next set of iterative training begins based on the newly selected test dataset and training dataset. After completing all iterative training rounds in a batch, the trained parameter model is validated using the local validation dataset, and parameter models with validation accuracy greater than or equal to a specified accuracy threshold are added to the model cache pool and then enter the next batch of training.

[0009] The technical solution provided in this embodiment can achieve a relatively complete model training process by setting different batches, number of groups and number of rounds. Furthermore, by setting a global validation set, a local validation dataset, a test dataset and a training dataset during the training process, the efficiency and accuracy of the training process can be guaranteed.

[0010] In one embodiment, the method further includes: if the verification accuracy is less than a specified accuracy threshold, discarding the parameter model trained in the current batch and proceeding to the training process of the next batch.

[0011] In one implementation, the first ratio is 5%, the second ratio is 5%, the third ratio is 10%, the fourth ratio is 80%, the specified batch is 100 times, the specified number of groups is 15 groups, and the specified number of rounds is 30 rounds.

[0012] In one implementation, the weight coefficients of each feature parameter group are determined as follows: for any feature parameter group, the number of parameter models in the model cache pool of the feature parameter group is identified, and the training batch of the parameter models of the feature parameter group during the training process is determined; the ratio between the number of parameter models and the training batch is calculated, and the ratio is used as the model fitting parameter of the feature parameter group; the proportion of the model fitting parameter of the feature parameter group in the total model fitting parameters of each feature parameter group is calculated, and the proportion is used as the weight coefficient of the feature parameter group.

[0013] The technical solution provided in this embodiment utilizes the number of parameter models and training batches in the model cache pool to obtain relatively accurate model fitting parameters, thereby obtaining relatively accurate weight coefficients, which provides an accurate data foundation for generating the final risk prediction level.

[0014] In one implementation, during the training and validation of the parameter model using the data sample set, the parameter model processes the data samples in the data sample set as follows: if the feature parameter group contains only one factor parameter, the data samples are normalized, and a three-layer fully connected network is used to classify the normalized features. Each fully connected layer incorporates L2 regularization for correction and an early stopping mechanism to prevent overfitting. Regarding the loss function, a custom loss function is defined based on the ReLU function, which penalizes cases where the actual sample is in an unhealthy state but the predicted result is healthy. If the feature parameter group contains more than one factor parameter... The weights of the parameter model are split into two parts: the first part of the split weights enters the random forgetting branch, and the second part of the split weights enters the feature processing branch, which includes convolutional layers, batch normalization layers, and pooling layers. The weights output by the random forgetting branch and the feature processing branch are integrated and then enter the attention layer for dynamic weight allocation. A three-layer fully connected network is used to classify the features after dynamic weight allocation. In each fully connected network layer, an L2 regularization mechanism is added for correction, and an early stopping mechanism is added to prevent overfitting. Regarding the loss function, a custom loss function is defined based on the ReLU function. This loss function adds a penalty for cases where the actual sample is in an unhealthy state, but the prediction result is in a healthy state.

[0015] The technical solution provided in this embodiment allows for the use of parameter models with different structures for different numbers of factor parameters. When the feature parameter group contains only a single factor parameter, a three-layer fully connected network can be used to process the input sample and obtain the corresponding prediction result. When the feature parameter group contains multiple factor parameters, since factor parameter fusion is involved, the model's receptive field can be gradually increased by first splitting the weights and then integrating them, allowing it to pay more attention to the overall situation of the sample and enabling the model to make better risk predictions based on multiple factor parameters. Subsequently, dynamic weight allocation through an attention layer can obtain more accurate fused features, and based on these accurate fused features, an accurate risk prediction level can be obtained.

[0016] In one implementation, selecting representative models for each feature parameter group from the model cache pool includes: for any feature parameter group, validating the parameter model corresponding to the feature parameter group in the model cache pool using a pre-set global validation set; if the parameter model correctly predicts the validation data representing the idle stop state in the global validation set, determining the parameter model as a candidate parameter model; obtaining the validation result of the candidate parameter model for the global validation set; if the validation result meets the preset validation conditions, generating the cumulative residual of the candidate parameter model based on the validation result, and determining the candidate parameter model with the smallest cumulative residual as the representative model of the feature parameter group.

[0017] In one implementation, the verification result satisfying the preset verification conditions includes: for each verification sample in the global verification set, the risk type represented by the verification result is consistent with the pre-labeled risk type, the risk type including having risk and no risk; and the error between the risk prediction score represented by the verification result and the pre-labeled risk standard score is within a specified range.

[0018] The technical solution provided in this embodiment uses a global validation set to validate the parameter models in the model cache pool, which can filter out representative models that meet the requirements. These representative models can accurately predict the risk type of the validation samples. At the same time, they can also meet the error requirements for more detailed risk prediction scores, thereby achieving high prediction accuracy.

[0019] In one embodiment, the method further includes: if the score of the risk prediction level characterization of the target engine exceeds a preset score threshold within a specified time period, disassembling the high-pressure turbine blades in the target engine and obtaining parameter data of the disassembled high-pressure turbine blades; and adding the obtained parameter data to the global validation set and the training dataset respectively.

[0020] The technical solution provided in this embodiment indicates that if a target engine has a high risk prediction score, there is a risk of in-flight shutdown. In this case, the target engine can be disassembled, and the parameter data of the high-pressure turbine blades obtained from the disassembly can be fed back into the HPTB full life cycle database, thereby providing data support for the training dataset and the global validation set, which is beneficial for training a more accurate parameter model in the future.

[0021] Secondly, this application also provides a life prediction device for high-pressure turbine blades in aero-engines, the device comprising:

[0022] The feature decomposition unit is used to statistically analyze multidimensional factor parameters related to aero-engines and decompose the multidimensional factor parameters into multiple feature parameter groups. The multidimensional factor parameters are used to characterize engine performance parameters and / or flight status parameters. Each feature parameter group includes one or more factor parameters.

[0023] The model adaptation unit is used to obtain a data sample set of the feature parameter group for any feature parameter group, and select a parameter model that is adapted to the feature parameter group according to the number of factor parameters contained in the feature parameter group; wherein, if the number of factor parameters contained in the feature parameter group is more than 1, the selected parameter model shall include at least a parallel network branch for performing weight splitting.

[0024] The model training unit is used to train and validate the parameter model using the data sample set, and add the parameter model whose validation results meet the preset conditions to the model cache pool.

[0025] The prediction unit is used to select representative models for each feature parameter group from the model cache pool, and use the selected representative models to perform multi-dimensional predictions on the actual data of the target engine to obtain multiple single-dimensional prediction results.

[0026] A risk assessment unit is used to perform a weighted summation of the multiple single-dimensional prediction results based on the weight coefficients assigned to each of the feature parameter groups, so as to generate a risk prediction level for the target engine, wherein the risk prediction level is used to characterize the service life of the high-pressure turbine blades in the target engine.

[0027] Thirdly, this application also provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements the above-mentioned method for predicting the life of high-pressure turbine blades in aero-engines. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 A step diagram illustrating a method for predicting the lifespan of high-pressure turbine blades in an aero-engine, provided as an embodiment of this application;

[0030] Figure 2 This is a schematic diagram of the life prediction process for high-pressure turbine blades in an aero-engine according to one embodiment of this application.

[0031] Figure 3 This is a flowchart illustrating the specific processing of the parameter model in one application scenario of this application;

[0032] Figure 4 A schematic diagram of the functional modules of a life prediction device for high-pressure turbine blades in an aero-engine provided in one embodiment of this application;

[0033] Figure 5 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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, 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.

[0035] Furthermore, the use of terms such as "first," "second," etc., in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of embodiments in this application, unless otherwise stated, "multiple" means two or more. Additionally, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0036] During aircraft operation, all data generated by sensors is recorded in a specific format, known as QAR (Quick Access Recorder) data. This data is automatically uploaded to the cloud after the aircraft lands. Engineers can use QAR data to reconstruct flight status, engine operating conditions, and other information, which can serve as a basis for fault diagnosis. QAR data is recorded at different frequencies, increasing with the importance of the sensor whose parameters are being recorded, typically once per second.

[0037] Currently, airlines manage flight data based on aircraft tail numbers. In practice, components (including engines) are sometimes removed and installed on other aircraft. When tracking and managing these components is required, data becomes chaotic and difficult to trace completely. This is clearly undesirable for component management.

[0038] To address this situation, this application proposes a component-based full lifecycle database. This management method centers on the component and establishes a cloud-based electronic database. From the first installation of the component, it records information such as installation time, component serial number, and installation location. Each time related work is performed on the component, it supplements the record with information such as removal time, reason for removal, and work performed, in order to facilitate data management and traceability of the entire lifecycle of the HPTB engine model.

[0039] In practical applications, this application provides a feature extraction method for QAR data of aircraft engines. This method can quickly extract the accumulated operating conditions of engines from massive amounts of historical QAR data. Specifically, different labels can be used to correspond to different risk levels, and each risk level can correspond to different score intervals. The correspondence between labels and score intervals can be used for subsequent model training. In a specific application example, the score intervals corresponding to different labels are shown in Table 1.

[0040] Table 1 Score ranges for different labels

[0041]

[0042] In this embodiment, the risk level of the Label can correspond to the severity of HPTB crack failure. Specifically, the risk level can be as follows:

[0043] Level 7: The HPTB breaks into at least two parts along the small neck, and the turbine disk cannot prevent the part above the small neck from detaching from the turbine disk. The broken part can completely detach from the engine (fly out).

[0044] Level 6: HPTB breaks into two parts along the small neck, but the remaining material is still restrained by the turbine disk and will not come out;

[0045] Grade 5: HPTB has a crack extending from the neck to the end face and the crack extends to a third face (e.g., the initiation face, the end face, the bottom face, or another side face).

[0046] Grade 4: HPTB has a crack extending from the neck to the end face and extending more than 1 mm into the end face, but does not extend to a third face (e.g., the initiation face, end face, bottom face, or another side face) at the bottom.

[0047] Grade 3: HPTB has cracks extending to the end face along the neck, but the extension length within the end face is less than 1 mm;

[0048] Level 2: HPTB has cracks in the pressure surface area, and FPI indicates that the crack length is greater than 2 mm, but it does not extend to the end face;

[0049] Level 1: HPTB has microcracks in the pressure surface area, and the FPI indicates that the crack length is less than 2 mm, which is defined as the crack initiation stage;

[0050] Grade 0: HPTB shows no cracks in the pressure surface area.

[0051] Among them, level 1 is the crack initiation stage, levels 1-3 are the crack growth stage, levels 3-4 are the rapid development stage, levels 4-6 are the dangerous outbreak stage, and level 7 is the failure stage.

[0052] Based on the HPTB status when the engine is removed, HPTB at level 2 or below (including level 2) can be defined as healthy, while HPTB at level 3 or above (including level 3) can be defined as unhealthy.

[0053] Please see Figure 1 and Figure 2 One embodiment of this application provides a method for predicting the lifespan of high-pressure turbine blades in an aero-engine, which may include the following steps.

[0054] S1: Collect multidimensional factor parameters related to aero-engines and decompose the multidimensional factor parameters into multiple characteristic parameter groups. The multidimensional factor parameters are used to characterize engine performance parameters and / or flight status parameters. Each characteristic parameter group includes one or more factor parameters.

[0055] In this embodiment, the multidimensional factor parameters related to the aero-engine can be engine performance parameters or flight state parameters. For example, in practical applications, factor parameters can be engine performance parameters such as N2 rotational speed and EGT (Exhaust Gas Temperature), or flight state parameters such as climb rate and climb time. These factor parameters can respectively characterize the different states experienced by the aero-engine. In this application, the multidimensional factor parameters can be decomposed into multiple feature parameter groups, and then a parameter model can be trained for each feature parameter group.

[0056] Specifically, the decomposed feature parameter set may include one or more factor parameters. For example, the N2 rotational speed can be a feature parameter set on its own, and the N2 parameter and the engine vibration parameter can also be combined into a feature parameter set. In other words, different feature parameter sets may contain different numbers of factor parameters. By training the model independently for each feature parameter set, the failure mechanism can be made explicit to a certain extent, laying the foundation for subsequent life prediction.

[0057] S2: For any set of feature parameters, obtain the data sample set of the feature parameter set, and select a parameter model that is suitable for the feature parameter set according to the number of factor parameters contained in the feature parameter set; wherein, if the number of factor parameters contained in the feature parameter set is more than 1, the selected parameter model shall include at least a parallel network branch for performing weight splitting.

[0058] In this implementation, an independent parameter model can be trained for each set of feature parameters. Before training, a data sample set for the feature parameter set can be obtained. Specifically, the data samples in the data sample set should be consistent with the type of the feature parameter set. For example, if the feature parameter set contains only a single climb rate, then the data samples should also contain climb rates. Or, for example, if the feature parameter set contains both climb rate and climb time as factor parameters, then the data samples should also contain both climb rate and climb time as factor parameters. Each data sample can have its own label, which can be used to represent the true risk score.

[0059] In this embodiment, different parameter models with different structures can be selected for the feature parameter set based on the number of factor parameters included in the feature parameter set. Please refer to [link / reference]. Figure 3 If the feature parameter set contains only a single factor parameter, then the selected parameter model can be executed. Figure 3The data processing flow is shown on the left side. Specifically, if the feature parameter group contains only one factor parameter, the data samples can be normalized first, and a three-layer fully connected network can be used to classify the normalized features. Each fully connected layer incorporates L2 regularization for correction and an early stopping mechanism to prevent overfitting. Regarding the loss function, a custom loss function is defined based on the ReLU function. This loss function penalizes cases where the actual sample is unhealthy but the predicted result is healthy, ensuring the model can better capture faults. Specifically, when calculating the loss residual, a correction factor α (α>1) is introduced when the actual sample is unhealthy but the predicted score is in an interval two or more intervals lower than the actual sample. By multiplying the loss residual by α, the result focuses more on correctly predicting unhealthy actual samples. Here, α is a variable that increases with the increase of the loss residual.

[0060] If the feature parameter set contains more than one factor parameter, then the selected parameter model can be executed. Figure 3 The processing procedure is shown on the right side. Specifically, if the feature parameter group contains more than one factor parameter, after normalization, the weights of the parameter model can be split. The first part of the split weights enters the random forgetting branch, and the second part enters the feature processing branch, which includes convolutional layers, batch normalization layers, and pooling layers. The weights output by the random forgetting branch and the feature processing branch are integrated and then enter the attention layer for dynamic weight allocation. In practical applications, the weight splitting and weight integration process can be repeated three times. After dynamic weight allocation in the attention layer, similar to the processing of a single factor parameter, a three-layer fully connected network is used to classify the dynamically weighted features. Each fully connected layer incorporates L2 regularization for correction and an early stopping mechanism to prevent overfitting. Regarding the loss function, a custom loss function is defined based on the ReLU function, which penalizes cases where the actual sample is in an unhealthy state but the predicted result is healthy. The weight splitting process can be viewed as dividing the input data into two parts: x1 and x2. One part goes through a convolutional layer, a batch normalization layer, and a pooling layer to obtain F(x2), while the other part goes through a random forgetting branch to obtain x1'. The final output is y = x1' + F(x2), thus ensuring that gradient vanishing and exploding do not occur.

[0061] The technical solution provided in this embodiment allows for the use of parameter models with different structures for different numbers of factor parameters. When the feature parameter group contains only a single factor parameter, a three-layer fully connected network can be used to process the input sample and obtain the corresponding prediction result. When the feature parameter group contains multiple factor parameters, since factor parameter fusion is involved, weight splitting can be performed first, followed by weight integration. This process can be performed three times, with the convolution kernel gradually increasing and the receptive field gradually increasing, enabling the model to pay more attention to the overall situation of the sample and to make better risk predictions based on multiple factor parameters. Subsequently, dynamic weight allocation through the attention layer can obtain more accurate fused features, and based on these accurate fused features, an accurate risk prediction level can be obtained.

[0062] S3: Use the data sample set to train and validate the parameter model, and add the parameter model whose validation results meet the preset conditions to the model cache pool.

[0063] In this embodiment, after obtaining the data sample set and selecting the corresponding parameter model for the feature parameter group, the parameter model can be trained and validated using the data sample set.

[0064] In practical applications, training processes with different batches, different numbers of groups, and different rounds can be set up. Specifically, a first proportion of data is extracted from the data sample set as a global validation set, which includes data samples representing healthy states and data samples representing unhealthy states.

[0065] The parameter model is trained in specified batches. Each batch of training includes a specified number of iterative training sessions, and each iterative training session includes a specified number of training rounds. For any batch of training, a second proportion of data is selected from the data sample set other than the global validation set as a local validation dataset. In any iterative training session, a third proportion of data is selected as the test dataset, and a fourth proportion of data is selected as the training dataset.

[0066] In each round of training, the parameter model is trained using the training dataset, and after training is completed, the trained parameter model is tested using the test dataset to determine the gradient descent direction.

[0067] After completing all training and testing rounds in a set of iterative training, a new test dataset and training dataset are selected, and the next set of iterative training begins based on the newly selected test dataset and training dataset.

[0068] After completing iterative training for all groups in a batch, the trained parametric models are validated using the local validation dataset. Parametric models with validation accuracy greater than or equal to a specified accuracy threshold are added to the model cache pool and proceed to the next batch of training. If the validation accuracy is less than the specified accuracy threshold, the parametric models trained in the current batch are discarded, and the next batch of training begins.

[0069] In a specific application example, the first ratio is 5%, the second ratio is 5%, the third ratio is 10%, the fourth ratio is 80%, the specified batch size is 100 times, the specified number of groups is 15 groups, and the specified number of rounds is 30 rounds. The above training process can be described as follows:

[0070] The initial work involved selecting 5% of the engines from the data sample set, without participating in any training, and only performing global accuracy verification as the aforementioned global verification set. In the global verification set, the number of healthy and unhealthy data samples was roughly equal.

[0071] During the first batch of training, 5% of the engines covered by the data sample set (not duplicated from the global validation set) are selected as the local validation dataset for the first batch. The local validation dataset is not used for training.

[0072] In the first training phase, data from 10% of the engines covered by the data sample set was selected as the test dataset, while data from the remaining 80% of the engines was used as the training dataset. The training dataset was fed into the parametric model for training. After each epoch, the parametric model was tested using the test dataset to determine the gradient descent direction, for a total of 30 training epochs. After 30 epochs, the selection of the test and training datasets was repeated (the proportions remained the same, but the data was re-selected), and the next set of training was started, for a total of 15 training sets.

[0073] After completing the above process, i.e., after 15*30 rounds of training, validation is performed. The trained parametric model is validated using a selected local validation dataset. If the validation accuracy is greater than or equal to 0.8 (the validation result meets the preset conditions), the trained parametric model is saved to the model cache pool; otherwise, the trained parametric model is not saved. After one batch of training is completed, regardless of whether the parametric model is saved to the model cache pool, the next batch of training is performed.

[0074] Specifically, the verification accuracy can be calculated using the following formula:

[0075] TPVariance=(label i -predict i)2

[0076]

[0077] Where TPVariance represents the error variance, label i The value predicted represents the true risk score of the i-th validation sample. i Let represent the predicted risk score of the i-th validation sample, acc represent the validation accuracy, and n represent the total number of validation samples in the global validation set.

[0078] The technical solution provided in this embodiment can achieve a relatively complete model training process by setting different batches, number of groups and number of rounds. Furthermore, by setting a global validation set, a local validation dataset, a test dataset and a training dataset during the training process, the efficiency and accuracy of the training process can be guaranteed.

[0079] In one implementation, after training the models for each feature parameter group is complete, weight coefficients can be set for each feature parameter group based on the number of parameter models retained in the model cache pool. Specifically, for any feature parameter group, the number of parameter models for that feature parameter group in the model cache pool is identified, and the training batch of the parameter models for that feature parameter group during the training process is determined. The ratio between the number of parameter models and the training batch is calculated, and this ratio is used as the model fitting parameter for that feature parameter group. The proportion of the model fitting parameter for that feature parameter group in the total model fitting parameters of all feature parameter groups is calculated, and this proportion is used as the weight coefficient for that feature parameter group.

[0080] For example, after training with one hundred batches for the N2 rotational speed dimension:

[0081] The model cache pool stores k models in the N2 rotational speed dimension, trained in a total of m batches (here, 100). Therefore, the model fitting parameters R in the N2 rotational speed dimension are... i It can be represented as:

[0082]

[0083] After all feature parameter sets have been trained, the total model fitting parameters R can be expressed as:

[0084] R = ∑R

[0085] Then the weighting coefficient w corresponding to the N2 rotational speed dimension i Represented as:

[0086]

[0087] Ultimately, any one of w iIt should be a number between 0 and 1, w i The larger the value, the easier it is for the model to converge, and the larger the corresponding weight; w i The smaller the value, the more difficult it is to correctly classify the result based on the current set of feature parameters. Therefore, models with a validation accuracy greater than 0.8 are likely accidental phenomena caused by overfitting, and their corresponding weights are smaller.

[0088] The technical solution provided in this embodiment utilizes the number of parameter models and training batches in the model cache pool to obtain relatively accurate model fitting parameters, thereby obtaining relatively accurate weight coefficients, which provides an accurate data foundation for generating the final risk prediction level.

[0089] S4: Select representative models for each feature parameter group from the model cache pool, and use the selected representative models to perform multi-dimensional predictions on the actual data of the target engine to obtain multiple single-dimensional prediction results.

[0090] In this embodiment, after training all batches is completed, the model cache pool can store the parametric models corresponding to each feature parameter group. Typically, because each feature parameter group has undergone training in a specified number of batches, the model cache pool contains a relatively large number of parametric models corresponding to that feature parameter group. In practical applications, when using parametric models for risk prediction, representative models for each feature parameter group can be selected, and these representative models can be used to perform risk prediction for their respective dimensions.

[0091] Specifically, for any set of feature parameters, a pre-set global validation set can be used to validate the parameter model corresponding to the set of feature parameters in the model cache pool. If the parameter model correctly predicts the validation data representing the idle stop state in the global validation set, the parameter model is determined as a candidate parameter model.

[0092] Then, the verification results of the candidate parameter model for the global verification set can be obtained. If the verification results meet the preset verification conditions, the cumulative residual of the candidate parameter model can be generated according to the verification results, and the candidate parameter model with the smallest cumulative residual can be determined as the representative model of the feature parameter group.

[0093] In practical applications, the verification result satisfying the preset verification conditions includes the following two processes:

[0094] For each validation sample in the global validation set, the risk type represented by the validation result is consistent with the pre-labeled risk type, which includes both risk and no risk; and

[0095] The error between the risk prediction score represented by the verification result and the pre-labeled risk standard score is within a specified range.

[0096] For example, in this application, the risk levels can be pre-classified. All risk levels can be divided into risky and risk-free. Risky can correspond to level 3 or above (including level 3) in Table 1, and risk-free can correspond to level 2 or below (including level 2) in Table 1. In other words, risky can correspond to HPTB defined as unhealthy, and risk-free can correspond to HPTB defined as healthy.

[0097] The first requirement is that, under this coarse classification, the prediction score output by the candidate parameter model should not be lower than 0.7 (the threshold for level 3) for all risky samples in the global validation set, meaning that all risky samples are correctly predicted as risky samples. Similarly, risk-free samples also need to be correctly predicted as risk-free samples.

[0098] The second requirement is that, assuming the first requirement is met, if the predicted risk score deviates significantly from the standard risk score, it indicates insufficient accuracy of the candidate parameter model. Therefore, an error range can be set for the predicted risk score. Only models within this range can be adopted as the final representative model. Specifically, according to the range setting, the risk score predicted by the candidate parameter model should not fluctuate by more than one level from the standard risk score. For example, if the actual risk level is 6, a predicted risk score range of 5-7 is acceptable. However, this fluctuation must meet the first requirement. For instance, if the actual risk level is 3, a predicted risk level of 2 is unacceptable (because a risky sample is incorrectly predicted as a risk-free sample).

[0099] When both of the above conditions are met, the cumulative residual of the model for each candidate parameter can be calculated: (labelPoint - predictPoint), where n is the total number of validation samples in the global validation set, labelPoint is the true risk score, and predictPoint is the risk score predicted by the candidate parameter model. Finally, the candidate parameter model with the lowest cumulative residual can be selected as the representative model.

[0100] The technical solution provided in this embodiment uses a global validation set to validate the parameter models in the model cache pool, which can filter out representative models that meet the requirements. These representative models can accurately predict the risk type of the validation samples. At the same time, they can also meet the error requirements for more detailed risk prediction scores, thereby achieving high prediction accuracy.

[0101] In one embodiment, the method further includes: if the score of the risk prediction level characterization of the target engine exceeds a preset score threshold within a specified time period, disassembling the high-pressure turbine blades in the target engine and obtaining parameter data of the disassembled high-pressure turbine blades; and adding the obtained parameter data to the global validation set and the training dataset respectively.

[0102] The technical solution provided in this embodiment indicates that if a target engine has a high risk prediction score, there is a risk of in-flight shutdown. In this case, the target engine can be disassembled, and the parameter data of the high-pressure turbine blades obtained from the disassembly can be fed back into the HPTB full life cycle database, thereby providing data support for the training dataset and the global validation set, which is beneficial for training a more accurate parameter model in the future.

[0103] In this embodiment, after obtaining the representative models for each feature parameter group, these representative models can be used to perform multi-dimensional predictions on the actual data of the target engine, resulting in multiple single-dimensional prediction results. Of course, during the prediction process, each representative model only predicts the factor parameters contained within its own feature parameter group. For example, the representative model corresponding to N2 speed will only predict the N2 speed data of the target engine.

[0104] S5: Based on the weight coefficients assigned to each of the feature parameter groups, the multiple single-dimensional prediction results are weighted and summed to generate a risk prediction level for the target engine, wherein the risk prediction level is used to characterize the service life of the high-pressure turbine blades in the target engine.

[0105] In this embodiment, after obtaining multiple single-dimensional prediction results, the weight coefficients of each feature parameter group predetermined in the aforementioned steps are used to perform a weighted summation of these single-dimensional prediction results, thereby obtaining the final risk score. This final risk score can then be used to look up the corresponding risk prediction level of the target engine in Table 1. Each risk prediction level corresponds to the HPTB state, thus allowing the prediction of the HPTB's service life.

[0106] The technical solution provided in this embodiment decomposes multidimensional factor parameters related to aero-engines into multiple characteristic parameter groups. By training models on each characteristic parameter group, individual parameter models for each group can be obtained. During training, different parameter models with varying structures can be selected based on the number of factor parameters contained in each characteristic parameter group, making the feature processing more accurate. Subsequently, by selecting representative models for each characteristic parameter group, different single-dimensional prediction results can be obtained using these representative models. By weighted summing of these single-dimensional prediction results, an accurate risk prediction level can be obtained, which can be used to characterize the service life of high-pressure turbine blades. The technical solution provided in this embodiment enables accurate prediction of engine risk levels without engine disassembly, thereby determining the service life of high-pressure turbine blades and effectively avoiding in-flight engine shutdowns.

[0107] In one implementation, if the score representing the risk prediction level of the target engine exceeds a preset score threshold within a specified time period (the preset score threshold can be flexibly set according to the actual application scenario), the high-pressure turbine blades in the target engine can be disassembled, and the parameter data of the disassembled high-pressure turbine blades can be obtained. The parameter data can be added to the global validation set and the training dataset respectively.

[0108] The technical solution provided in this embodiment indicates that if a target engine has a high risk prediction score, there is a risk of in-flight shutdown. In this case, the target engine can be disassembled, and the parameter data of the high-pressure turbine blades obtained from the disassembly can be fed back into the HPTB full life cycle database, thereby providing data support for the training dataset and the global validation set, which is beneficial for training a more accurate parameter model in the future.

[0109] Please see Figure 4 One embodiment of this application also provides a life prediction device for high-pressure turbine blades in an aero-engine, the device comprising:

[0110] The feature decomposition unit is used to statistically analyze multidimensional factor parameters related to aero-engines and decompose the multidimensional factor parameters into multiple feature parameter groups. The multidimensional factor parameters are used to characterize engine performance parameters and / or flight status parameters. Each feature parameter group includes one or more factor parameters.

[0111] The model adaptation unit is used to obtain a data sample set of the feature parameter group for any feature parameter group, and select a parameter model that is adapted to the feature parameter group according to the number of factor parameters contained in the feature parameter group; wherein, if the number of factor parameters contained in the feature parameter group is more than 1, the selected parameter model shall include at least a parallel network branch for performing weight splitting.

[0112] The model training unit is used to train and validate the parameter model using the data sample set, and add the parameter model whose validation results meet the preset conditions to the model cache pool.

[0113] The prediction unit is used to select representative models for each feature parameter group from the model cache pool, and use the selected representative models to perform multi-dimensional predictions on the actual data of the target engine to obtain multiple single-dimensional prediction results.

[0114] A risk assessment unit is used to perform a weighted summation of the multiple single-dimensional prediction results based on the weight coefficients assigned to each of the feature parameter groups, so as to generate a risk prediction level for the target engine, wherein the risk prediction level is used to characterize the service life of the high-pressure turbine blades in the target engine.

[0115] The specific implementation methods of each unit module can be found in the description of the aforementioned method implementation method, and will not be repeated here.

[0116] Please see Figure 5 This application also provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and when the computer program is executed by the processor, it implements the above-described method for predicting the life of high-pressure turbine blades in aero-engines.

[0117] like Figure 5 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 5 Take a processor 10 as an example.

[0118] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0119] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0120] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0121] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0122] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0123] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the methods shown in the above embodiments are implemented.

[0124] This application provides a computer program product including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method of any embodiment of this application.

[0125] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

[0126] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0131] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0132] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0133] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for predicting the lifespan of high-pressure turbine blades in an aero-engine, characterized in that, The method includes: Statistically analyze multidimensional factor parameters related to aero-engines and decompose these multidimensional factor parameters into multiple characteristic parameter groups. The multidimensional factor parameters are used to characterize engine performance parameters and / or flight status parameters. Each characteristic parameter group includes one or more factor parameters. For any set of feature parameters, obtain the data sample set of the feature parameter set, and select a parameter model that is suitable for the feature parameter set according to the number of factor parameters contained in the feature parameter set; wherein, if the number of factor parameters contained in the feature parameter set is more than 1, the selected parameter model shall include at least a parallel network branch for performing weight splitting. The parameter model is trained and validated using the data sample set, and the parameter model whose validation results meet the preset conditions is added to the model cache pool; Representative models are selected for each feature parameter group from the model cache pool, and the selected representative models are used to make multi-dimensional predictions on the actual data of the target engine, resulting in multiple single-dimensional prediction results. Based on the weight coefficients assigned to each of the feature parameter groups, the multiple single-dimensional prediction results are weighted and summed to generate a risk prediction level for the target engine, wherein the risk prediction level is used to characterize the service life of the high-pressure turbine blades in the target engine. During the training and validation of the parametric model using the data sample set, the parametric model processes the data samples in the data sample set in the following manner: If the feature parameter group contains only one factor parameter, the data sample is normalized, and a three-layer fully connected network is used to classify the normalized features. In each fully connected network layer, an L2 regularization mechanism is added for correction, and an early stopping mechanism is added to prevent overfitting. A loss function is customized based on the ReLU function. The loss function adds a penalty for cases where the actual sample is in an unhealthy state, but the prediction result is in a healthy state. If the feature parameter group contains more than one factor parameter, the weights of the parameter model are split. The first part of the split weights enters the random forgetting branch, and the second part of the split weights enters the feature processing branch. The feature processing branch includes convolutional layers, batch normalization layers, and pooling layers. The weights output by the random forgetting branch and the feature processing branch are integrated and then enter the attention layer for dynamic weight allocation. A three-layer fully connected network is used to classify the features after dynamic weight allocation. An L2 regularization mechanism is added to each fully connected network layer for correction, and an early stopping mechanism is added to prevent overfitting. A custom loss function is defined based on the ReLU function. The loss function adds a penalty for cases where the actual sample is in an unhealthy state, but the prediction result is in a healthy state.

2. The method according to claim 1, characterized in that, Training and validating the parameter model using the data sample set, and adding parameter models whose validation results meet preset conditions to the model cache pool includes: A first proportion of data is extracted from the data sample set as a global validation set, the global validation set including data samples representing healthy status and data samples representing unhealthy status; The parameter model is trained in specified batches. Each batch of training includes a specified number of iterative training sets, and each iterative training set includes a specified number of training rounds. For any batch of training, a second proportion of data is selected from the data sample set other than the global validation set as a local validation dataset. In any iterative training set, a third proportion of data is selected as the test dataset, and a fourth proportion of data is selected as the training dataset. In each round of training, the parameter model is trained using the training dataset, and after training is completed, the trained parameter model is tested using the test dataset to determine the gradient descent direction. After completing all training and testing rounds in a set of iterative training, a new test dataset and training dataset are selected, and the next set of iterative training begins based on the newly selected test dataset and training dataset. After completing iterative training of all groups in a batch, the trained parameter models are validated using the local validation dataset. Parameter models with validation accuracy greater than or equal to a specified accuracy threshold are added to the model cache pool and then proceed to the next batch of training.

3. The method according to claim 2, characterized in that, The method further includes: If the verification accuracy is less than the specified accuracy threshold, the parameter model obtained from the current batch of training will be discarded, and the training process will begin for the next batch.

4. The method according to claim 2, characterized in that, The first ratio is 5%, the second ratio is 5%, the third ratio is 10%, the fourth ratio is 80%, the specified batch is 100 times, the specified number of groups is 15 groups, and the specified number of rounds is 30 rounds.

5. The method according to claim 1 or 2, characterized in that, The weight coefficients of each of the aforementioned feature parameter groups are determined in the following manner: For any set of feature parameters, identify the number of parameter models that the set of feature parameters has in the model cache pool, and determine the training batch of the parameter models of the set of feature parameters during the training process. Calculate the ratio between the number of parameter models and the training batch, and use the ratio as the model fitting parameter for the feature parameter group; Calculate the proportion of the model fitting parameters of the feature parameter group in the total model fitting parameters of each feature parameter group, and use the proportion as the weight coefficient of the feature parameter group.

6. The method according to claim 1, characterized in that, Selecting representative models for each feature parameter group from the model cache pool includes: For any set of feature parameters, the parameter model corresponding to the set of feature parameters in the model cache pool is verified using a pre-set global verification set. If the parameter model correctly predicts the verification data representing the idle stop state in the global verification set, the parameter model is determined as a candidate parameter model. Obtain the verification results of the candidate parameter model for the global verification set. If the verification results meet the preset verification conditions, generate the cumulative residual of the candidate parameter model based on the verification results, and determine the candidate parameter model with the smallest cumulative residual as the representative model of the feature parameter group.

7. The method according to claim 6, characterized in that, The verification result satisfies the preset verification conditions, including: For each validation sample in the global validation set, the risk type represented by the validation result is consistent with the pre-labeled risk type, which includes both risk and no risk; and The error between the risk prediction score represented by the verification result and the pre-labeled risk standard score is within a specified range.

8. The method according to claim 1, characterized in that, The method further includes: If the score representing the risk prediction level of the target engine exceeds a preset score threshold within a specified time period, the high-pressure turbine blades in the target engine are disassembled, and the parameter data of the disassembled high-pressure turbine blades are obtained. The acquired parameter data is added to the global validation set and the training dataset, respectively.

9. A life prediction device for high-pressure turbine blades in an aero-engine, characterized in that, The device includes: The feature decomposition unit is used to statistically analyze multidimensional factor parameters related to aero-engines and decompose the multidimensional factor parameters into multiple feature parameter groups. The multidimensional factor parameters are used to characterize engine performance parameters and / or flight status parameters. Each feature parameter group includes one or more factor parameters. The model adaptation unit is used to obtain a data sample set of the feature parameter group for any feature parameter group, and select a parameter model that is adapted to the feature parameter group according to the number of factor parameters contained in the feature parameter group; wherein, if the number of factor parameters contained in the feature parameter group is more than 1, the selected parameter model shall include at least a parallel network branch for performing weight splitting. The model training unit is used to train and validate the parameter model using the data sample set, and add the parameter model whose validation results meet the preset conditions to the model cache pool. The prediction unit is used to select representative models for each feature parameter group from the model cache pool, and use the selected representative models to perform multi-dimensional predictions on the actual data of the target engine to obtain multiple single-dimensional prediction results. A risk assessment unit is used to perform a weighted summation of the multiple single-dimensional prediction results based on the weight coefficients assigned to each of the feature parameter groups, so as to generate a risk prediction level for the target engine, wherein the risk prediction level is used to characterize the service life of the high-pressure turbine blades in the target engine. During the training and validation of the parametric model using the data sample set, the parametric model processes the data samples in the data sample set in the following manner: If the feature parameter group contains only one factor parameter, the data sample is normalized, and a three-layer fully connected network is used to classify the normalized features. In each fully connected network layer, an L2 regularization mechanism is added for correction, and an early stopping mechanism is added to prevent overfitting. A loss function is customized based on the ReLU function. The loss function adds a penalty for cases where the actual sample is in an unhealthy state, but the prediction result is in a healthy state. If the feature parameter group contains more than one factor parameter, the weights of the parameter model are split. The first part of the split weights enters the random forgetting branch, and the second part of the split weights enters the feature processing branch. The feature processing branch includes convolutional layers, batch normalization layers, and pooling layers. The weights output by the random forgetting branch and the feature processing branch are integrated and then enter the attention layer for dynamic weight allocation. A three-layer fully connected network is used to classify the features after dynamic weight allocation. An L2 regularization mechanism is added to each fully connected network layer for correction, and an early stopping mechanism is added to prevent overfitting. A custom loss function is defined based on the ReLU function. The loss function adds a penalty for cases where the actual sample is in an unhealthy state, but the prediction result is in a healthy state.

10. A computer device, characterized in that, The computer device includes a memory and a processor, the memory being used to store a computer program that, when executed by the processor, implements the method as described in any one of claims 1 to 8.

Images