An ES-based weight-adaptive method and system for weight optimization in aero-engine failure retrieval

Abstract

This invention provides a method and system for optimizing engine failure retrieval weights based on adaptive ES weights, belonging to the field of engine failure dataset retrieval and analysis technology. The method includes: S1, acquiring a preprocessed engine failure dataset and dividing and labeling it; S2, constructing an adaptive ES weight neural network model, using momentum and adaptive learning rate to improve learning and prediction efficiency; S3, training and optimizing the adaptive ES weight neural network model using the engine failure dataset; and S4, fusing and integrating the adaptive ES weight neural network model using a weighted average method to output engine failure retrieval results. This invention constructs an adaptive ES weight neural network model, dynamically learning and adjusting the ES weights of fields through an adaptive neural network, making the retrieval results more closely resemble real-world engine failure scenarios, improving retrieval accuracy, and enabling rapid acquisition of relevant engine failure data information for diagnosing the causes of engine failures.

Description

Technical Field

[0001] This invention relates to the field of engine fault dataset retrieval and analysis technology, specifically to an aero-engine fault retrieval weight optimization method and system based on ES weight adaptive. Background Technology

[0002] In the era of information explosion, how to quickly and accurately retrieve the required information from massive amounts of data has become a key issue in the field of information processing. ES, or Elasticsearch, is an open-source distributed search engine based on the full-text search engine Lucene. With its powerful full-text search and real-time analysis capabilities, it occupies an important position in the data processing field. It can efficiently process petabytes of data, whether structured or unstructured, and index and retrieve both. Its distributed architecture, high availability, and scalability provide enterprises with stable and reliable search services, and it is widely used in record management, search engine construction, data analysis and processing, and many other fields. Adaptive neural networks, with their self-learning, self-organizing, and adaptive capabilities, can automatically adjust the network's weights and structure based on the characteristics of the input data. Introducing adaptive neural networks into ES weight adjustment allows for dynamic learning and adjustment of the weights of various fields based on different search requests and data distributions, thereby more accurately reflecting search intent and improving the relevance and quality of search results. Furthermore, adaptive adjustment of ES weights helps optimize the performance of the search system, reduce unnecessary waste of computing resources, and improve search efficiency, which has significant practical implications for search applications handling large-scale data.

[0003] Numerous studies have explored ES weighting and related optimizations. Some research improves the traditional TF-IDF algorithm by combining it with machine learning techniques to optimize ES weight calculation and enhance the relevance of search results. For different data types, such as images, text, and audio, research explores how to customize ES weight settings to adapt to specific data retrieval needs. Considering the characteristics of the Chinese language and business scenarios, a series of methods for optimizing ES weights have been proposed, such as combining Chinese word segmentation technology and semantic understanding to more accurately allocate weights to text fields.

[0004] In summary, there are still some shortcomings in applying adaptive neural networks to ES weight adjustment. On the one hand, most studies only involve adjusting ES weights based on a single factor, such as relying solely on text content and lacking comprehensive utilization of multi-source information. On the other hand, existing models have limited adaptability when facing dynamically changing data and complex retrieval requirements, making it difficult to quickly and accurately adjust ES weights to adapt to new situations. Furthermore, there are also shortcomings in model efficiency and interpretability, which limits their widespread application in real-world production environments. Therefore, there is still considerable room for further research into how to more effectively utilize adaptive neural networks to optimize ES weights to improve information retrieval performance and quality. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a method and system for optimizing engine failure retrieval weights based on adaptive ES weights. By dynamically learning and adjusting the ES weights of fields through an adaptive neural network, the retrieval results better reflect the needs of real-world engine failure scenarios, significantly improving retrieval accuracy and recall. This allows for the rapid acquisition of relevant engine failure data, aiding in the diagnosis of engine failure causes. The invention constructs an adaptive neural network model based on ES weights, which automatically adapts to different retrieval scenarios and data characteristics. Based on multi-source information, this adaptive neural network model comprehensively mines the potential value within engine failure datasets, more accurately determining ES weights and enhancing the relevance of retrieval results. A dynamic adaptive algorithm based on real-time feedback enables the model to adjust ES weights promptly based on retrieval results, quickly adapting to dynamic changes in engine failure datasets and the needs of engine failure analysis scenarios, achieving dynamic adaptation and improving the flexibility and versatility of the retrieval system.

[0006] Specifically, on the one hand, the present invention provides a weight optimization method for aero-engine failure retrieval based on ES weight adaptation, which includes the following steps:

[0007] S1: Obtain the exhaust temperature, vibration amplitude spectrum, lubricating oil metal shavings content, compressor pressure ratio and flow fluctuation value, and abnormal fuel flow of the aero-engine to form an engine fault dataset. Then, perform preprocessing, divide and label the engine fault dataset for training an adaptive neural network model.

[0008] S2: Construct an adaptive neural network model for ES weights, introduce the adaptive neural network model into the ES weight adjustment, and generate the ES weights for each engine fault dataset field through the adaptive neural network model; use momentum method and adaptive learning rate to improve the learning and prediction efficiency of the adaptive neural network model;

[0009] S3: Using the engine exhaust temperature, lubricating oil metal shavings content, and abnormal fuel flow rate from the engine fault dataset obtained in step S1, an engine fault data vector is constructed. This vector is fused with word vectors and input into the ES weight adaptive neural network model constructed in step S2 for training. A regularized loss function is constructed to adjust the process parameters, and the ES weight adaptive neural network model is optimized. Specifically:

[0010] ;

[0011] ;

[0012] in, To perform regularized optimization of the loss function for predicting engine fault datasets based on ES weights; Optimize the loss function for prediction based on the ES-weighted engine fault dataset; For regularization functions; The number of samples in the engine fault dataset; Number the samples in the engine fault dataset; For the first The actual expected output of a sample of an engine fault dataset; For the first Predicted output for a sample of an engine fault dataset; The regularization coefficient is used. For the ES weight adaptive neural network model, the first Each ES weight;

[0013] S4: The ES weighted adaptive neural network model optimized in step S3 is integrated using a weighted average method to output the engine failure retrieval results.

[0014] Preferably, step S2 specifically includes:

[0015] S21: Integrate an adaptive neural network into the ES weight adjustment process to automatically learn and adjust the ES weights of the engine fault dataset, achieve dynamic optimization of ES weights, and construct an adaptive neural network model for ES weights.

[0016] S22: Optimize the network structure of the ES weight adaptive neural network model; adopt a multi-layer feedforward neural network structure, including: input layer, hidden layer and output layer, which are connected by ES weights to realize the transmission and processing of engine fault dataset;

[0017] S23: By combining momentum and adaptive learning rate with gradient descent, the weights of the adaptive neural network model ES are adjusted through the calculation of the loss function, thereby improving the relevance and accuracy of engine fault dataset retrieval results.

[0018] Preferably, the input layer, hidden layer, and output layer in step S2 are specifically as follows:

[0019] The input layer receives an external engine fault dataset and performs word segmentation; it uses a pre-trained engine fault dataset word vector model to obtain the word vector representation of the engine fault dataset, which also corresponds to the number of neurons in the input layer.

[0020] The hidden layers extract and transform features from the input engine fault dataset; the hidden layers are connected by a weight matrix and nonlinear transformation is performed using the ReLU activation function to improve the training efficiency and performance of the ES weight adaptive neural network model.

[0021] The number of neurons in the output layer is the same as the number of fields in the engine fault dataset. It receives engine fault data information from the hidden layer and outputs the ES weight value of each field in the engine fault dataset after linear transformation.

[0022] Preferably, the momentum method in step S23 specifically includes:

[0023] The formula for updating the weights of the engine fault dataset ES using the momentum method is:

[0024] ;

[0025] ;

[0026] in, for Engine fault dataset ES weight momentum at time +1; for Engine fault dataset ES weighted momentum at time step; The momentum factor for the engine fault dataset ES; for The weights of the engine fault dataset ES at each time step; for The weights of the engine fault dataset ES at each time step; This is the learning rate coefficient; For loss function exist Time-based weights of the engine fault dataset ES The gradient; This is a time parameter.

[0027] Preferably, the adaptive learning rate in step S23 is as follows:

[0028] The adaptive learning rate dynamically adjusts the learning rate based on changes in the loss function during training, specifically as follows:

[0029] ;

[0030] in, for The learning rate coefficient at any given time; This is the first attenuation parameter; This is the second attenuation parameter.

[0031] Preferably, step S3 specifically includes:

[0032] S31: Input the engine fault dataset training data, update the ES weights through forward propagation, calculate the loss, and backpropagation to complete the training of the ES weight adaptive neural network model;

[0033] S32: Based on the feedback from the engine fault dataset validation set, construct a loss function and use learning rate adjustment and regularization coefficient adjustment to optimize and adjust the parameters of the ES weight adaptive neural network model to improve the performance of the ES weight adaptive neural network model.

[0034] S33: Optimize the ES weight adaptive neural network model using L1 regularization, L2 regularization, and Dropout techniques respectively to improve the model's generalization and prevent overfitting.

[0035] Preferably, the loss function in step S31 is as follows:

[0036] To measure the difference between the model's predicted output and the actual expected output, using an engine fault dataset as input, the loss function of the ES weighted adaptive neural network model is optimized, resulting in:

[0037] ;

[0038] in, Optimize the loss function for prediction based on the ES-weighted engine fault dataset.

[0039] Preferably, in step S33, the L1 regularization and L2 regularization methods are used to update the ES weights, specifically as follows:

[0040] ;

[0041] ;

[0042] in, This represents the gradient of the original loss function with respect to the weights. ES weights The symbolic function.

[0043] Preferably, step S4 specifically includes:

[0044] Training was completed using step S3. There are three independent ES weight adaptive neural network models, and the prediction results of each ES weight adaptive neural network model are as follows: The final fusion prediction engine failure retrieval results obtained through the weighted average method The formula is:

[0045] ;

[0046] in, The results of the engine failure search; For the first Output of an ES weighted adaptive neural network model; The total number of ES weight adaptive neural network models; This refers to the ES weight adaptive neural network model number; To determine the ES weight adaptive neural network model Key parameters of contribution; This is a sample of the input engine fault dataset.

[0047] On the other hand, the present invention provides an aero-engine fault data retrieval system based on an ES weight adaptive aero-engine fault retrieval weight optimization method, which includes: an engine fault dataset acquisition module, an ES weight adaptive neural network model construction module, an ES weight adaptive neural network model optimization module, and an ES weight adaptive neural network model fusion and integration module.

[0048] The engine fault dataset acquisition module constructs an engine fault dataset based on the ES index, enabling the adaptive neural network model to learn comprehensive engine fault dataset ES weights, thereby improving analysis and retrieval efficiency and accuracy; it preprocesses the engine fault dataset to improve its quality; and it performs engine fault dataset partitioning and labeling.

[0049] The ES weight adaptive neural network model building module integrates adaptive neural networks into the ES weight adjustment process. It utilizes the self-learning and dynamic adjustment efficiency of adaptive neural networks to automatically learn and adjust the ES weights of fields, thereby improving the relevance and accuracy of retrieval results. It uses momentum method and adaptive learning rate combined with gradient descent method to adjust the ES weights of the adaptive neural network model by calculating the loss function, thereby improving the relevance and accuracy of engine fault dataset retrieval results.

[0050] The ES weighted adaptive neural network model optimization module trains the ES weighted adaptive neural network model, adjusts the ES weighted adaptive neural network model parameters, and is the key to improving the performance of the ES weighted adaptive neural network model. Optimizing the adjustment of the ES weighted adaptive neural network model parameters improves the model's generalization and prevents overfitting.

[0051] The ES weighted adaptive neural network model fusion and integration module uses a weighted average method to fuse and train N independent ES weighted adaptive neural network models, and outputs engine failure retrieval results.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] (1) This invention can improve the retrieval accuracy of engine fault data information; by dynamically learning and adjusting the field ES weight through adaptive neural network, the retrieval results are more in line with the needs of real aero-engine failure scenarios, significantly improving the retrieval accuracy and recall rate, and can quickly obtain relevant engine fault data information to help diagnose the cause of aero-engine failure.

[0054] (2) The present invention constructs an adaptive neural network model based on ES weights, which significantly enhances the adaptability of the model compared with the traditional model; the model can automatically adapt to different retrieval scenarios and data features without the need for frequent manual intervention to adjust the ES weights, thereby improving the flexibility and versatility of the retrieval system.

[0055] (3) The present invention proposes an adaptive neural network model based on multi-source information, which can fully explore the potential value in the engine fault dataset, more accurately determine the ES weight, and improve the relevance of the retrieval results; the dynamic adaptive algorithm based on real-time feedback enables the model to adjust the ES weight in a timely manner according to the retrieval results, quickly adapt to the dynamic changes in the engine fault dataset and the requirements of the aero-engine failure analysis scenario, and achieve dynamic adaptation. Attached Figure Description

[0056] Figure 1 The control block diagram is shown for the weight optimization method for aero-engine failure retrieval based on ES weight adaptive.

[0057] Figure 2 This is a flowchart illustrating the engine failure retrieval process according to an embodiment of the present invention. Detailed Implementation

[0058] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0059] This invention proposes a weight optimization method for aero-engine failure retrieval based on ES weight adaptation, such as... Figure 1 As shown, the process involves obtaining a preprocessed engine fault dataset and dividing and labeling it; constructing an ES weighted adaptive neural network model, using momentum and adaptive learning rate to improve learning and prediction efficiency; training and optimizing the ES weighted adaptive neural network model using the engine fault dataset; and fusion and integrating the ES weighted adaptive neural network model using a weighted average method to output engine failure retrieval results. The specific steps include:

[0060] Step S1: Obtain the engine failure dataset, preprocess it, and filter out the aero-engine failure data related to engine failure for training the adaptive neural network model.

[0061] Step S11: Construct an engine fault dataset based on the ES index. Extract engine fault datasets from engine fault dataset reports, covering physical quantities such as exhaust temperature, vibration amplitude and spectrum, lubricating oil metal shavings content, compressor pressure ratio and flow fluctuations, and abnormal fuel flow. These physical quantities are interrelated, and their dynamic changes collectively reveal the engine's health status and early signs of failure, enabling the adaptive neural network model to learn engine fault ES weights and improve analysis and retrieval efficiency and accuracy. The engine fault dataset is collected using the ES application programming interface (API), and corresponding queries are written to retrieve engine fault datasets containing different engine fault causes from the ES index.

[0062] In the field of engine fault analysis and retrieval, the Elasticsearch (ES) retrieval API is used to query physical quantities and corresponding descriptions of engine faults, including exhaust temperature, lubricating oil metal shavings content, and abnormal fuel flow. This physical quantity and fault description data serves as the foundation for training an adaptive neural network. Simultaneously, to improve the timeliness and accuracy of the engine fault dataset, the fault data is updated regularly. In the engine fault dataset record analysis scenario, ES's record collection function is used to acquire engine fault datasets containing fields such as timestamp, record level, and record content. These are then used to train an adaptive neural network model to meet the needs of engine fault dataset record retrieval.

[0063] Step S12: Preprocess the engine fault dataset. Because the engine fault dataset collected in step S11 has issues such as noise, missing values, and inconsistent formats, preprocessing is necessary to improve its quality and provide a reliable foundation for training the adaptive neural network model.

[0064] Step S121: Perform engine fault dataset cleaning and preprocessing. In this embodiment of the invention, tools such as regular expressions and data dictionaries are used to identify and correct errors and outliers in the engine fault dataset.

[0065] Step S122: Perform word segmentation preprocessing on the engine fault dataset, dividing the dataset into individual engine fault terms. For Chinese word segmentation of the engine fault dataset, the IK segmentation tool is used in this embodiment to accurately segment the Chinese text into engine fault dataset terms. When processing the title and body text fields of the aero-engine fault dataset, the Jieba segmentation tool is used to segment the text describing the engine fault dataset for subsequent extraction of text features. For English word segmentation of the engine fault dataset, commonly used tools include Natural Language Toolkit (NLTK), which can segment words according to English grammar and vocabulary rules. For the English title and abstract fields of academic literature, NLTK is used for word segmentation, providing a foundation for subsequent text analysis.

[0066] Step S123: Perform preprocessing and labeling of the engine fault dataset; assigning labels or category information to the engine fault dataset is crucial for supervised learning training. In the field of engine fault dataset retrieval, labeling mainly determines the relevance between the engine fault dataset and the retrieval query, thereby improving the retrieval efficiency of the engine fault dataset.

[0067] Step S13: Divide and label the engine fault dataset. Divide the preprocessed engine fault dataset from Step S12 into a training set, a validation set, and a test set to improve the training, evaluation, and generalization capabilities of the adaptive neural network model. Typically, this is done according to a certain ratio. For example, in this embodiment, 70% of the engine fault dataset is allocated to the training set for training the adaptive neural network model; 15% is allocated to the validation set to monitor the performance of the adaptive neural network model during training, adjust its parameters, and prevent overfitting; and the remaining 15% is allocated to the test set to evaluate the final performance of the trained adaptive neural network model.

[0068] Random sampling is used to partition the engine fault dataset, ensuring that the distribution of engine fault datasets in each subset is representative. In this embodiment of the invention, the `random` library in Python is used to randomly shuffle the preprocessed engine fault dataset, and then the dataset is divided into training, validation, and test sets according to the aforementioned proportions. A fixed random seed is set to ensure the stability and repeatability of the partitioning. Similarly, when partitioning the engine fault dataset, a random seed of 42 is set, and then random sampling is performed. This ensures that the partitioning results are consistent across multiple runs of the partitioning code.

[0069] For the annotation of engine fault datasets, supervised learning training primarily involves labeling each engine fault dataset with its relevance score to the retrieval query. The annotated engine fault datasets then serve as input for training an adaptive neural network. The adaptive neural network model learns from these annotated engine fault datasets and continuously adjusts the weights of the engine fault dataset ES (Extended Search Engine) to improve the relevance and accuracy of the retrieval results.

[0070] Step S2: Construct an ES weighted adaptive neural network model and use momentum method and adaptive learning rate to improve the learning and prediction efficiency of the adaptive neural network model on the engine fault dataset in step S1.

[0071] Step S21: Design an adaptive neural network model for ES weights. To overcome the limitations of traditional ES weight settings, this invention integrates an adaptive neural network into the ES weight adjustment process to achieve dynamic optimization of ES weights. Its core design idea is based on the ES retrieval process, utilizing the self-learning and dynamic adjustment efficiency of the adaptive neural network to automatically learn and adjust field ES weights according to different retrieval requests and data characteristics, thereby improving the relevance and accuracy of retrieval results.

[0072] In the Elasticsearch (ES) retrieval process, a search request is first received, parsed, and segmented. Then, relevant engine fault datasets are found using the inverted index, and a relevance score is calculated between the engine fault dataset and the search results. Finally, the search results are ranked based on the score. In this process, the ES weight settings directly affect the calculation of the relevance score, thus impacting the quality of the search results. Traditional fixed ES weights cannot be flexibly adjusted according to changes in search requests and data, resulting in poor relevance of the search results.

[0073] Adaptive neural network models can automatically adjust the ES weights and network structure based on the characteristics of the input engine fault dataset, exhibiting strong adaptability and learning capabilities. Introducing this into ES weight adjustment leverages this characteristic, using multi-source information such as engine fault retrieval requests and engine fault datasets as input. Through learning and training, the adaptive neural network model dynamically generates ES weights for each engine fault field. Specifically, when a retrieval request is received, keywords, semantic information, and historical search records from the request are used as input to the input layer, while the content of each field in the document is used as another input. The adaptive neural network model learns and analyzes this input information, performing feature extraction and transformation in the hidden layer to uncover latent patterns and relationships within the engine fault dataset. The output layer then outputs the ES weight values ​​for each field in the engine fault dataset. These ES weight values ​​are applied to the relevance score calculation, thereby achieving dynamic adjustment of ES weights according to different retrieval scenarios and the characteristics of the engine fault dataset, improving the matching degree between retrieval results and needs.

[0074] In the original search weights, the title field has a weight of 1, the time field has a weight of 1.5, and the engine failure cause field has a weight of 2. When the search results are retrieved, instead of clicking on the top-ranked engine failure dataset, the search terms and clicked on subsequent engine failure datasets are automatically recorded for training the ES weight adaptive neural network model. After training, the ES weight adaptive neural network model will readjust the weights of each field to obtain a more accurate ranking result.

[0075] Step S22: Optimize the network structure of the ES weight adaptive neural network model; adopt a multi-layer feedforward neural network structure, including: input layer, hidden layer and output layer, and the layers are connected by ES weights to realize the transmission and processing of information.

[0076] The input layer receives an external engine fault dataset, and the number of neurons is determined based on the number of features in the input engine fault dataset. In the ES weighted adaptive neural network model, the input engine fault dataset includes: retrieval request information and document field content information. For the retrieval request, after word segmentation, each engine fault data point is treated as a feature, corresponding to one neuron in the input layer. Simultaneously, to capture the semantic information of the retrieval request, pre-trained engine fault dataset word vector models can be used, such as Word2Vec (for generating engine fault dataset word vectors) or GloVe (a text representation technology model based on the ES weighted adaptive neural network model), to convert engine fault dataset words into engine fault dataset word vector representations. The dimension of these engine fault dataset word vectors also corresponds to the number of neurons in the input layer. For example, using a 300-dimensional engine fault dataset word vector model, the number of input layer neurons corresponding to each engine fault dataset word is 300. Furthermore, historical retrieval behavior, such as retrieval frequency and the field distribution of clicked documents, is also used as input engine fault data information, and the corresponding number of neurons is determined based on the number of its features. For the engine fault dataset field content, each field's content is also segmented and vectorized. The vector dimension corresponding to each field also corresponds to a neuron in the input layer. Assuming the engine fault dataset document contains three fields: title, description, and tag, with each field represented by a 200-dimensional vector, the number of neurons in this input layer is 200 × 3 = 600. Combining the engine fault dataset retrieval request and the field content information, the total number of neurons in the input layer is the sum of the numbers in both fields.

[0077] The hidden layers are the core of the ES weighted adaptive neural network model, responsible for feature extraction and transformation of the input engine fault dataset. The ES weighted adaptive neural network model uses multiple hidden layers to enhance its expressive power, enabling it to learn more complex patterns and relationships within the engine fault dataset. Each hidden layer contains a number of neurons, the number of which is adjusted based on the specific task and data complexity. Generally, the number of neurons in the hidden layers falls between the number of neurons in the input and output layers; for example, the first hidden layer might have 1.5 times the number of neurons in the input layer, with subsequent hidden layers gradually decreasing in number to achieve progressive abstraction and feature extraction of the engine fault dataset. The weight matrices between hidden layers are determined by the number of neurons in the previous and current layers. Within the hidden layers, the ReLU activation function is used to perform a non-linear transformation on the neuron inputs, introducing non-linearity and enabling the ES weighted adaptive neural network model to learn the non-linear relationships within the engine fault dataset. The ReLU activation function is expressed as f(x) = max(0,x). When the input engine fault dataset x is greater than 0, the output is the engine fault dataset x; when the input engine fault dataset x is less than or equal to 0, the output is 0. By using the ReLU function, the vanishing gradient problem can be effectively avoided, improving the training efficiency and performance of the ES weight adaptive neural network model.

[0078] The output layer has the same number of neurons as the number of fields in the engine failure dataset, with each neuron corresponding to the ES weight for each field in the engine failure dataset. The output layer receives information from the hidden layers, performs a linear transformation, and outputs the ES weight value for each field in the engine failure dataset. No activation function is used in the output layer; the ES weight values ​​are output directly to more intuitively reflect the importance of the fields. For example, for an engine failure dataset containing three fields: engine failure dataset title, engine failure dataset description, and engine failure dataset label, the output layer will output three weight values, corresponding to the ES weights of the engine failure dataset title, engine failure dataset description, and engine failure dataset label fields, respectively. These ES weight values ​​are applied to the ES weight relevance score calculation to adjust the importance of different fields in the relevance evaluation of search results.

[0079] Neurons in each layer are connected via ES weights, which are continuously adjusted during the training of the ES weight adaptive neural network model to optimize its performance. The initial values ​​of the ES weights are typically randomly initialized to avoid limiting the learning ability of the model due to identical initial ES weights. During training, the gradient of each ES weight is calculated based on the error between the predicted output and the actual expected output using the backpropagation algorithm. Then, the Adam optimization algorithm is used to update the ES weights based on the gradients, allowing the predicted output of the ES weight adaptive neural network model to gradually approach the actual expected output, thereby achieving accurate learning and adjustment of the ES weights.

[0080] Step S23: Adjust the ES weights of the adaptive neural network model by calculating the loss function using momentum and adaptive learning rate combined with gradient descent. To train the ES weight adaptive neural network model so that it can accurately learn and adjust the ES weights, an improved gradient descent method is used as the ES weight adjustment algorithm. This algorithm is based on the traditional gradient descent method, combined with momentum and adaptive learning rate adjustment strategies to improve the convergence speed and stability of the algorithm and avoid getting trapped in local optima. The traditional gradient descent method calculates the gradient of the loss function with respect to the ES weights, and then updates the ES weights in the opposite direction of the gradient to minimize the loss function; the ES weight update formula for the engine fault dataset is:

[0081] ;

[0082] in for The weights of the engine fault dataset ES at each time step; for The weights of the engine fault dataset ES at each time step; This is the learning rate coefficient; For loss function exist Time-based weights of the engine fault dataset ES The gradient; This is a time parameter.

[0083] However, traditional gradient descent methods have some drawbacks, such as slow convergence speed, susceptibility to local optima, and sensitivity to the choice of learning rate. To overcome these problems, this invention introduces the momentum method, which can accelerate weight updates to a certain extent and skip some local optima. The core idea of ​​the momentum method is to combine the current gradient with the previous ES weight update direction during weight updates, similar to the concept of momentum in physics; the ES weight update formula for the engine fault dataset is obtained as follows:

[0084] ;

[0085] ;

[0086] in, Let ES be the weighted momentum of the engine fault dataset at time t+1; Let be the weighted momentum of the engine fault dataset ES at time t; The weight momentum factor for the engine fault dataset ES is typically around 0.9.

[0087] Furthermore, to better adapt to the characteristics of engine fault datasets at different training stages, this algorithm also employs an adaptive learning rate adjustment strategy. A traditional fixed learning rate may lead to excessively large weight updates in the early stages of training, preventing convergence; conversely, it may lead to excessively small weight updates in the later stages, slowing down convergence. The adaptive learning rate adjustment strategy dynamically adjusts the learning rate based on changes in the loss function during training. For example, a learning rate decay strategy can be used, gradually decreasing the learning rate as the number of training epochs increases. The specific decay formula can be:

[0088] ;

[0089] in, for The learning rate coefficient at any given time; This is the first attenuation parameter; This is the second decay parameter; adjusting the decay parameter controls the decay rate of the learning rate.

[0090] In the actual training process, the predicted output of the ES weighted adaptive neural network model is first calculated based on the input engine fault dataset. Then, the loss function is calculated based on the predicted output and the actual expected output. The commonly used loss function is the mean squared error (MSE) loss function, and its calculation formula is as follows:

[0091] ;

[0092] in, The mean squared error (MSE) loss function; The number of samples in the engine fault dataset; For the first The actual expected output of a sample of an engine fault dataset; For the first The predicted output of a dataset of engine faults.

[0093] Next, the gradient of the loss function with respect to the ES weights is calculated using the error backpropagation algorithm. Starting from the output layer, the error is propagated back layer by layer to the input layer, calculating the contribution of each ES weight to the loss function, thus obtaining the gradient. Finally, based on the improved gradient descent formula, combined with momentum and adaptive learning rate adjustments, the ES weights of the adaptive neural network model are updated. Through continuous iterative training, the loss function gradually decreases, enabling the adaptive neural network model to learn accurate ES weights, thereby improving the relevance and accuracy of engine fault dataset retrieval results.

[0094] Step S3: Train the ES weight adaptive neural network model constructed in step S2 using the engine fault dataset, adjust the process parameters, and optimize the ES weight adaptive neural network model.

[0095] Step S31: Train the ES weight adaptive neural network model. The training process of the ES weight adaptive neural network model is an iterative optimization process, including key steps such as: inputting training data from the engine fault dataset, forward propagation, calculating the loss, and backpropagation to update the ES weights.

[0096] At the start of training, the pre-defined engine fault dataset and training set are input into the ES weighted adaptive neural network model. Using the exhaust temperature from the engine fault dataset, the exhaust temperature is divided into four stages: start-up, ground idling, takeoff, and cruise. Since the exhaust temperature range differs for each stage, and the corresponding outliers also differ, different numbers are used to represent each stage, for example, 0-3. The lubricating oil metal content uses the iron, copper, chromium, and aluminum content as its representative physical parameters. The abnormal fuel flow rate, like the exhaust temperature, is divided into four stages, and the corresponding stage's representation number and abnormal fuel flow rate are used as parameters. These three parameters form the engine fault data vector, which is fused with word vectors and input into the neural network. If the engine fault data vector contains only one physical quantity, the other parameters use 0 as placeholders in the input vector.

[0097] After receiving the engine fault dataset, the input layer propagates forward to the hidden layer through ES weight connections between neurons. In the hidden layer, neurons transform the input vector containing word vectors and physical quantities. Each neuron in the hidden layer is connected to the neurons in the previous layer via an ES weight matrix and undergoes a non-linear transformation using the ReLU activation function. For the input engine fault dataset... After passing through the ES weight matrix and bias After calculation, we get Then, through the activation function ReLU function The transformation is performed to obtain the output of the hidden layer. Through the layer-by-layer processing of multiple hidden layers, the features of the engine fault dataset are gradually extracted and abstracted, and the ES weighted adaptive neural network model is able to learn more complex patterns and relationships in the engine fault dataset.

[0098] The results processed by the hidden layer continue to propagate forward to the output layer. Based on the received information, the output layer calculates the ES weight value for each engine fault dataset field. The number of neurons in the output layer is the same as the number of fields in the document, and each neuron corresponds to the weight of each engine fault dataset field. The output layer converts the output of the hidden layer into ES weight values ​​through a linear transformation, directly outputting the weight values ​​without using an activation function.

[0099] Calculating the loss function is a crucial step in training an ES weighted adaptive neural network model. It measures the difference between the model's predicted output and the actual expected output. The mean squared error (MSE) function is typically used as the loss function, and its formula is as follows:

[0100] ;

[0101] in, The loss function is used to predict and optimize the engine fault dataset; n is the number of samples in the engine fault dataset. For the first The actual expected output of a sample of an engine fault dataset; For the first Predicted output for a sample of an engine fault dataset; The sample number is assigned to the engine fault dataset.

[0102] Since the engine fault dataset used as input data in the ES weighted adaptive neural network model contains considerable noise interference, the existing loss function formula is optimized to obtain:

[0103] ;

[0104] in, Optimize the loss function for prediction based on the ES-weighted engine fault dataset.

[0105] In the ES weight adaptive neural network model, the actual expected output is the true ES weight of each engine fault dataset field determined based on the labeled engine fault dataset, while the predicted output is the ES weight calculated by the ES weight adaptive neural network model. The loss function is optimized based on the ES weight engine fault dataset prediction. The value of the loss function is used to evaluate the prediction accuracy of the ES weight adaptive neural network model. The smaller the loss value, the closer the predicted output of the ES weight adaptive neural network model is to the actual expected output. Backpropagation is based on the prediction optimization loss function of the ES weight engine fault dataset. The value of each ES weight is calculated to determine its contribution to the loss, and the error is backpropagated to previous layers to update the ES weights of the ES weight adaptive neural network model. The backpropagation algorithm is based on the chain rule, starting from the output layer and propagating the error back layer by layer to the input layer. At the output layer, the prediction optimization loss function based on the ES weight engine fault dataset is calculated. The gradient of the ES weights in the output layer is calculated, and then this gradient is backpropagated to the hidden layer. In the hidden layer, based on the received gradient, the prediction optimization loss function based on the ES weights engine fault dataset is computed. The ES gradient of the hidden layer weights is calculated through repeated backpropagation. Finally, based on the obtained gradients, the prediction optimization loss function based on the engine fault dataset of ES weights in step S22 is updated. The ES weights are used; through continuous iterative training, the loss function for predicting engine fault datasets based on ES weights is optimized. Gradually reduce the weights to continuously improve the performance of the ES weight adaptive neural network model.

[0106] Step S32: Optimize and adjust the parameters of the ES weight adaptive neural network model. During training, adjusting the parameters of the ES weight adaptive neural network model based on feedback from the engine fault dataset validation set is crucial for improving its performance. Parameter optimization strategies include adjusting the learning rate and regularization coefficients.

[0107] The learning rate is a crucial parameter controlling the step size of ES weight updates in an adaptive ES weight neural network model, significantly impacting its convergence speed and performance. If the learning rate is too large, the model may skip the optimal solution during training, leading to convergence failure; conversely, if the learning rate is too small, the convergence speed will be slow, requiring more training time and computational resources. Therefore, the learning rate needs to be dynamically adjusted based on the performance feedback from the engine fault dataset validation set. In the early stages of training, a larger learning rate allows the model to quickly explore the parameter space; as training progresses, the learning rate gradually decreases, allowing for more precise adjustment of the ES weights to approach the optimal solution. A learning rate decay strategy is employed, gradually decreasing the learning rate coefficient as the number of training epochs increases. Common learning rate decay methods include exponential decay and step size decay; the formula for exponential decay is:

[0108] ;

[0109] in, for The learning rate coefficient at any given time; The initial learning rate coefficient; The attenuation rate; This represents the number of training rounds.

[0110] Regularization is an important means of preventing overfitting in ES weighted adaptive neural network models. By adding a regularization term to the loss function, the complexity of the ES weighted adaptive neural network model is constrained. Common regularization methods include L1 regularization and L2 regularization. L1 regularization adds the sum of the absolute values ​​of the weights as a regularization term to the loss function, while L2 regularization adds the sum of the squares of the weights as a regularization term. Taking L2 regularization as an example, the loss function is:

[0111] ;

[0112] in, To perform regularized optimization of the loss function for predicting engine fault datasets based on ES weights; For regularization functions, based on regularization coefficients Sure.

[0113] Regularization coefficient The magnitude of the regularization coefficient determines the degree of constraint on the complexity of the ES weight adaptive neural network model. If the regularization coefficient is too large, the ES weight adaptive neural network model will be too simple and may result in underfitting; if the regularization coefficient is too large... If the regularization coefficient is too small, the complexity of the ES weight adaptive neural network model will not be effectively constrained, potentially leading to overfitting. Therefore, it is necessary to select an appropriate regularization coefficient based on the performance of the engine fault dataset and validation set, using methods such as cross-validation. During training, the validation set is divided into multiple subsets, each using a different regularization coefficient. Training and validation were performed, and the regularization coefficient that performed best on the engine fault dataset validation set was selected. As the final parameter.

[0114] Step S33: Optimize the ES weight adaptive neural network model using L1 regularization and L2 regularization methods respectively. These are important techniques for improving the model's generalization and preventing overfitting. In the training of the ES weight adaptive neural network model, commonly used regularization methods include L1 regularization, L2 regularization, and Dropout, each of which achieves model optimization through unique principles and implementation methods.

[0115] Step S331: L1 regularization, also known as Lasso regularization, works by adding the L1 norm of the ES weight vector to the loss function, which is the sum of the absolute values ​​of all ES weights. In the training of the ES weight adaptive neural network model, the loss function... Originally, the loss function only measured the error between the model's predicted output and the actual output, such as the mean squared error (MSE). After introducing L1 regularization, the loss function becomes:

[0116] ;

[0117] in, Optimize the loss function for predicting L1-regularized engine fault datasets based on ES weights; The regularization coefficient is used. For the ES weight adaptive neural network model, the first Each ES weight; This represents the number of samples in the engine fault dataset.

[0118] The purpose of L1 regularization is to reduce the value of some ES weights to 0, thereby achieving feature selection. In text classification tasks on engine fault datasets, the ES weights of some words may become 0 after L1 regularization. This means that these words contribute little to the classification result and are automatically ignored by the ES weight adaptive neural network model, making the model simpler and reducing the risk of overfitting. When implementing L1 regularization, the gradient of the L1 regularization term with respect to the weights is usually calculated simultaneously during backpropagation gradient calculation. This gradient is then incorporated into the ES weight update formula when updating the weights. When using gradient descent to update the ES weights, the ES weight update formula becomes:

[0119] ;

[0120] in, This is the learning rate coefficient; This represents the gradient of the original loss function with respect to the weights. ES weights The symbolic function.

[0121] Step S332: L2 regularization, also known as Ridge Regression, adds the L2 norm of the ES weight vector to the loss function, which is the sum of squares of all ES weights; at this point, the loss function becomes:

[0122] ;

[0123] in, The loss function is optimized for prediction of L2-regularized ES-weighted engine fault datasets.

[0124] The principle of L2 regularization is to penalize larger ES weights, causing them to tend towards smaller values. This makes the ES weight adaptive neural network model smoother and improves its generalization ability. In image recognition tasks, L2 regularization prevents the ES weight adaptive neural network model from overfitting to some noisy features in the engine fault dataset, allowing it to better capture the essential features of images in the engine fault dataset and thus exhibit better performance on the engine fault dataset test set. When implementing L2 regularization, the gradient of the L2 regularization term with respect to the ES weights is also calculated during backpropagation, and the ES weight update formula becomes:

[0125] ;

[0126] Step S333: Optimizing the ES weight adaptive neural network model using Dropout is a simple and effective regularization technique that prevents overfitting by randomly dropping a subset of neurons during training. Specifically, in each training iteration, some neurons are randomly selected with a certain probability, and their outputs are set to 0, as if these neurons were temporarily removed from the ES weight adaptive neural network model. In a multilayer perceptron model, some neurons in the hidden layers may be randomly dropped during training. This forces the ES weight adaptive neural network model to learn more robust feature representations because each neuron cannot depend on other specific neurons, thus reducing co-adaptation among neurons. Dropout is typically applied during forward propagation. During training, a mask equal to the number of neurons is generated based on the set dropout probability. The mask elements are either 0 or 1, where 0 indicates the neuron is dropped and 1 indicates the neuron is retained. During testing, Dropout is no longer used; instead, all neurons are used, but their outputs are multiplied by the retention probability from training to maintain consistency between the expected outputs of the ES weight adaptive neural network model during training and testing.

[0127] Step S4: The optimized ES weighted adaptive neural network model from Step S3 is fused using a weighted average method to output the engine failure retrieval result. Fusing or integrating multiple trained ES weighted adaptive neural network models is an effective way to further improve their performance. Common fusion methods include weighted average and voting methods. By combining the prediction results of multiple ES weighted adaptive neural network models, the accuracy and stability of the ES weighted adaptive neural network model are improved. The weighted average method, with its ability to differentiate the contribution of ES weighted adaptive neural network models, exhibits superior flexibility and accuracy in continuous value prediction scenarios, such as parameter prediction in regression tasks and trend estimation of time series data. It is particularly suitable for model ensemble scenarios with dynamic weight adjustment characteristics, such as ES weighted adaptive neural network models. The essence of the weighted average method is to assign different trust weights to each ES weighted adaptive neural network model based on its performance on the validation set. The final prediction result is the weighted sum of the predicted values ​​of each ES weighted adaptive neural network model and their corresponding weights. The core logic of this method is to give higher decision weight to the better-performing and more generalized ES weighted adaptive neural network model in the fusion result, while weakening the interference of the weaker ES weighted adaptive neural network model, thereby achieving a superior ensemble effect. Training is completed using step S3. An independent ES weight adaptive neural network model, denoted as... For a specific input engine fault dataset sample The prediction results of each ES weight adaptive neural network model are as follows: Assume each ES weight is an adaptive neural network model. Corresponding weight The final fusion prediction engine failure retrieval result obtained by the weighted average method is then... The formula is:

[0128] ;

[0129] in, The results of the engine failure search; For the first Output of an ES weighted adaptive neural network model; The total number of ES weight adaptive neural network models; This refers to the ES weight adaptive neural network model number; To determine the ES weight adaptive neural network model The key parameter of contribution directly affects the fusion effect. The larger the weight, the stronger the influence of the model's prediction result on the final output. This is a sample of the input engine fault dataset.

[0130] like Figure 2 The diagram shows the engine failure retrieval flowchart of an embodiment of the present invention. The experimental design of this embodiment is as follows: A comparative experiment was set up to comprehensively evaluate the performance of the field weight adjustment method of the ES weight adaptive neural network model. A series of comparative experiments were designed. The ES weight adaptive neural network model method is referred to as ANN-Weight, and compared with the traditional fixed field weight method, referred to as Fixed-Weight. In the Fixed-Weight method, the weights of each field are pre-set based on experience. For example, in the engine failure dataset scenario, the weight of the title field is set to 1, the weight of the aircraft engine failure cause field is set to 3, and the weights of the aircraft engine failure time, location, etc., are set to 2, etc., and these weights remain unchanged throughout the search process. A comparison was also made with an advanced weight adjustment method based on machine learning—the logistic regression weight adjustment method, referred to as LR-Weight. The LR-Weight method learns from historical search data and user feedback, and uses a logistic regression model to train and obtain the weights of each field in the engine failure dataset. In specific implementation, data such as search requests, document field content, and user click behavior on search results are used as features, and the relevance score between the document and the search query is used as a label. The logistic regression model is trained, and the field weights are determined by the coefficients of the model. To further validate the effectiveness of the ANN-Weight method, it was compared with a weight adjustment method based on a combination of rules and statistics, referred to as RS-Weight. The RS-Weight method first formulates basic weight rules for some common search scenarios based on domain knowledge and business rules. Then, it fine-tunes the weights by incorporating statistical characteristics of the data, such as term frequency and document frequency. In the news search scenario, the title field is assigned a higher weight according to the rules, and then the weights of the corresponding fields are appropriately adjusted based on the frequency of different keywords in the document. In the comparative experiments, different methods were tested under the same dataset and search query conditions to improve the fairness and comparability of the experiments. Each method was repeatedly tested under different experimental conditions, and the average value was taken as the final experimental result to reduce experimental error and improve the accuracy of the results.

[0131] To accurately measure the performance of different methods in search tasks, precision, recall, F1 score, and mean average precision (MAP) were selected as evaluation criteria.

[0132] Accuracy is used to measure the proportion of truly relevant documents among the retrieved documents. The formula is: Accuracy = Number of relevant documents retrieved / Total number of documents retrieved. In e-commerce search, if 100 product documents are retrieved, and 80 of them are truly relevant to the user's needs, then the accuracy rate is 80 / 100 = 0.8. A higher accuracy rate indicates fewer falsely identified documents as relevant, and thus higher search result accuracy.

[0133] Recall measures the ratio of relevant documents retrieved to the total number of relevant documents. The formula is: Recall = Number of relevant documents retrieved / Total number of relevant documents. In the e-commerce search example above, if there are 1000 relevant product documents and 80 are retrieved, the recall rate is 80 / 1000 = 0.08. A higher recall rate indicates that more relevant documents are found, and the more comprehensive the search is.

[0134] The F1 score is a metric for precision and recall. It is the harmonic mean of precision and recall, calculated as: F1 = 2 * Precision * Recall / (Precision + Recall). The F1 score provides a more comprehensive reflection of a search method's performance, avoiding biased evaluations that rely solely on precision or recall. In the previous search example, the F1 score = 2 * 0.8 * 0.08 / (0.8 + 0.08) ≈ 0.145. A higher F1 score indicates a better balance between accuracy and comprehensiveness in the search method.

[0135] Mean Precision (MAP) is used to evaluate the average performance of a retrieval across multiple queries. It is a weighted average of precision at different recall levels, and MAP is the average AP of all queries. MAP provides a more comprehensive reflection of retrieval performance under different query conditions; a higher MAP value indicates better overall performance. In actual MAP calculation, for each query, its precision at different recall levels is calculated. Then, the precision is weighted and summed based on the changes in recall to obtain the AP for that query. Finally, the APs of all queries are averaged to obtain the MAP.

[0136] These evaluation criteria reflect the performance of search methods from different perspectives. By comprehensively analyzing these indicators, we can fully and accurately evaluate the advantages and disadvantages of the field weight adjustment method of the ES weight adaptive neural network model in improving retrieval results.

[0137] The second aspect of this invention proposes an engine fault dataset analysis system based on an ES weight adaptive aero-engine fault retrieval weight optimization method, which includes: an engine fault dataset acquisition module, an ES weight adaptive neural network model construction module, an ES weight adaptive neural network model optimization module, and an ES weight adaptive neural network model fusion and integration module.

[0138] The engine fault dataset acquisition module constructs an engine fault dataset based on the ES index, enabling the adaptive neural network model to learn comprehensive engine fault dataset ES weights, thereby improving analysis and retrieval efficiency and accuracy. It preprocesses the engine fault dataset to improve its quality, providing a reliable foundation for training the adaptive neural network model. Finally, it partitions and labels the engine fault dataset into training, validation, and test sets to enhance the training, evaluation, and generalization capabilities of the adaptive neural network model.

[0139] The ES weight adaptive neural network model building module integrates adaptive neural networks into the ES weight adjustment process. Based on the ES retrieval process, it utilizes the self-learning and dynamic adjustment efficiency of adaptive neural networks to automatically learn and adjust field ES weights according to different retrieval requests and data characteristics, thereby improving the relevance and accuracy of retrieval results. It employs a multi-layer feedforward neural network structure, including an input layer, hidden layers, and an output layer, establishing connections between layers through ES weights to achieve information transmission and processing. Using momentum and adaptive learning rate combined with gradient descent, it calculates the loss function to adjust the ES weights of the adaptive neural network model, improving the relevance and accuracy of engine fault dataset retrieval results.

[0140] The ES weight adaptive neural network model optimization module trains the ES weight adaptive neural network model and adjusts the ES weight adaptive neural network model parameters based on feedback from the engine fault dataset validation set. This is the key to improving the performance of the ES weight adaptive neural network model. The ES weight adaptive neural network model parameters are optimized and adjusted. L1 regularization and L2 regularization methods are used to optimize the ES weight adaptive neural network model to improve the model's generalization and prevent overfitting.

[0141] The ES weighted adaptive neural network model fusion and integration module completes the fusion and integration training using the weighted average method. An independent ES weighted adaptive neural network model outputs engine failure retrieval results.

[0142] The beneficial effects of this invention are as follows: The embodiments of this invention construct an adaptive neural network model based on ES weights, which significantly enhances the model's adaptability compared to traditional models. By dynamically learning and adjusting the ES weights of fields through the adaptive neural network, the retrieval results better match the needs of real-world aero-engine failure scenarios, significantly improving retrieval accuracy and enabling rapid acquisition of relevant engine fault data information to help diagnose the causes of aero-engine failures. Based on multi-source information, the adaptive neural network model can comprehensively mine the potential value in engine fault datasets, more accurately determine ES weights, and improve the relevance of retrieval results. Based on a real-time feedback dynamic adaptive algorithm, the model can adjust ES weights in a timely manner according to the retrieval results, quickly adapting to the dynamic changes in engine fault datasets and aero-engine failure analysis scenario requirements, achieving dynamic adaptation, and meeting the needs of actual aero-engine fault diagnosis processes.

[0143] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An aero-engine failure retrieval weight optimization method based on ES weight self-adaption, characterized in that: It includes: S1: Obtain the exhaust temperature, lubricating oil metal shavings content, abnormal fuel flow, vibration amplitude spectrum, and compressor pressure ratio and flow fluctuation values ​​of the aero-engine to form an engine fault dataset. Then, perform preprocessing, divide and label the engine fault dataset for training an adaptive neural network model. S2: Construct an adaptive neural network model for ES weights, introduce the adaptive neural network model into the ES weight adjustment, and generate the ES weights for each engine fault dataset field through the adaptive neural network model; use momentum method and adaptive learning rate to improve the learning and prediction efficiency of the adaptive neural network model; S3: Using the engine exhaust temperature, lubricating oil metal shavings content, and abnormal fuel flow rate from the engine fault dataset obtained in step S1, an engine fault data vector is constructed. This vector is fused with word vectors and input into the ES weight adaptive neural network model constructed in step S2 for training. A regularized loss function is constructed to adjust the process parameters, and the ES weight adaptive neural network model is optimized. Specifically: ； ； in, To perform regularized optimization of the loss function for predicting engine fault datasets based on ES weights; Optimize the loss function for prediction based on the ES-weighted engine fault dataset; For regularization functions; The number of samples in the engine fault dataset; Number the samples in the engine fault dataset; For the first The actual expected output of a sample of an engine fault dataset; For the first Predicted output for a sample of an engine fault dataset; The regularization coefficient is used. For the ES weight adaptive neural network model, the first Each ES weight; S4: The ES weighted adaptive neural network model optimized in step S3 is integrated using a weighted average method to output the engine failure retrieval results.

2. The method for optimizing the weights of aero-engine failure retrieval based on ES weight adaptation according to claim 1, characterized in that: Step S2 is as follows: S21: Integrate an adaptive neural network into the ES weight adjustment process to automatically learn and adjust the ES weights of the engine fault dataset, achieve dynamic optimization of ES weights, and construct an adaptive neural network model for ES weights. S22: Optimize the network structure of the ES weight adaptive neural network model; A multi-layer feedforward neural network structure is adopted, including an input layer, a hidden layer and an output layer, which are connected by ES weights to realize the transmission and processing of engine fault datasets; S23: Using momentum and adaptive learning rate combined with gradient descent, the weights of the adaptive neural network model ES are adjusted by calculating the loss function.

3. The weight optimization method for aero-engine failure retrieval based on ES weight adaptation according to claim 2, characterized in that: The input layer, hidden layer, and output layer in step S22 are specifically as follows: The input layer receives an external engine fault dataset and performs word segmentation; it uses a pre-trained engine fault dataset word vector model to obtain the word vector representation of the engine fault dataset, which also corresponds to the number of neurons in the input layer. The hidden layers extract and transform features from the input engine fault dataset; the hidden layers are connected by a weight matrix and nonlinear transformation is performed using the ReLU activation function to improve the training efficiency and performance of the ES weight adaptive neural network model. The number of neurons in the output layer is the same as the number of fields in the engine fault dataset. It receives engine fault data information from the hidden layer and outputs the ES weight value of each field in the engine fault dataset after linear transformation.

4. The weight optimization method for aero-engine failure retrieval based on ES weight adaptation according to claim 2, characterized in that: The momentum method in step S23 is as follows: The formula for updating the weights of the engine fault dataset ES using the momentum method is: ； ； in, for Engine fault dataset ES weight momentum at time +1; for Engine fault dataset ES weighted momentum at time step; The momentum factor for the engine fault dataset ES; for The weights of the engine fault dataset ES at each time step; for The weights of the engine fault dataset ES at each time step; This is the learning rate coefficient; loss function exist Time-based weights of the engine fault dataset ES The gradient; This is a time parameter.

5. The weight optimization method for aero-engine failure retrieval based on ES weight adaptation according to claim 2, characterized in that: The adaptive learning rate in step S23 is as follows: The adaptive learning rate dynamically adjusts the learning rate based on changes in the loss function during training, specifically as follows: ； in, for The learning rate coefficient at any given time; This is the first attenuation parameter; This is the second attenuation parameter.

6. The weight optimization method for aero-engine failure retrieval based on ES weight adaptation according to claim 4, characterized in that: Step S3 is as follows: S31: Input the engine fault dataset training data, update the ES weights through forward propagation, calculate the loss, and backpropagation to complete the training of the ES weight adaptive neural network model; S32: Based on the feedback from the engine fault dataset validation set, construct a loss function and use learning rate adjustment and regularization coefficient adjustment to optimize and adjust the parameters of the ES weight adaptive neural network model. S33: Optimize the ES weight adaptive neural network model using L1 regularization, L2 regularization, and Dropout techniques respectively.

7. The method for optimizing the weights of aero-engine failure retrieval based on ES weight adaptation according to claim 6, characterized in that: The loss function in step S31 is as follows: To measure the difference between the model's predicted output and the actual expected output, using an engine fault dataset as input, the loss function of the ES weighted adaptive neural network model is optimized, resulting in: ； in, Optimize the loss function for prediction based on the ES-weighted engine fault dataset.

8. The weight optimization method for aero-engine failure retrieval based on ES weight adaptation according to claim 6, characterized in that: In step S33, the L1 regularization and L2 regularization methods implement ES weight updates, specifically as follows: ； ； in, This represents the gradient of the original loss function with respect to the weights. ES weights The symbolic function.

9. The method for optimizing the weights of aero-engine failure retrieval based on ES weight adaptation according to claim 1, characterized in that: Step S4 is as follows: Training was completed using step S3. There are three independent ES weight adaptive neural network models, and the prediction results of each ES weight adaptive neural network model are as follows: The final fusion prediction engine failure retrieval results obtained through the weighted average method The formula is: ； in, The results of the engine failure search; For the first Output of an ES weighted adaptive neural network model; The total number of ES weight adaptive neural network models; This refers to the ES weight adaptive neural network model number; To determine the ES weight adaptive neural network model Key parameters of contribution; This is a sample of the input engine fault dataset.

10. An aero-engine fault data retrieval system for the aero-engine fault retrieval weight optimization method based on ES weight adaptation as described in any one of claims 1 to 9, characterized in that, It includes: The module includes an engine fault dataset acquisition module, an ES weight adaptive neural network model construction module, an ES weight adaptive neural network model optimization module, and an ES weight adaptive neural network model fusion and integration module. The engine fault dataset acquisition module constructs an engine fault dataset based on the ES index, enabling the adaptive neural network model to learn comprehensive engine fault dataset ES weights, thereby improving analysis and retrieval efficiency and accuracy. Preprocess engine fault datasets to improve their quality; Perform engine fault dataset partitioning and annotation; The ES weight adaptive neural network model building module integrates adaptive neural networks into the ES weight adjustment process. It utilizes the self-learning and dynamic adjustment efficiency of adaptive neural networks to automatically learn and adjust the field ES weights, thereby improving the relevance and accuracy of retrieval results. It uses momentum method and adaptive learning rate combined with gradient descent method to adjust the ES weights of the adaptive neural network model by calculating the loss function. The ES weighted adaptive neural network model optimization module trains the ES weighted adaptive neural network model, adjusts the ES weighted adaptive neural network model parameters, and is the key to improving the performance of the ES weighted adaptive neural network model. Optimizing and adjusting the ES weighted adaptive neural network model parameters is crucial. The ES weighted adaptive neural network model fusion and integration module uses a weighted average method to fuse and train N independent ES weighted adaptive neural network models, and outputs engine failure retrieval results.

Images