Fault diagnosis method based on long and short term memory neural network

Abstract

The invention discloses a fault diagnosis method based on a long and short term memory neural network, and the method comprises the following steps: 1, collecting the related information of aviation equipment, and carrying out the data preprocessing, the data preprocessing content comprising data collection, data cleaning and data normalization; 2, feature extraction: performing feature extraction on the preprocessed data; 3, model construction: constructing an LSTM model according to the extracted features; 4, model training: training the constructed LSTM model; step 5, fault diagnosis is carried out, model testing is carried out, and a model testing result is obtained; and step 6, carrying out model optimization, optimizing the LSTM model according to a test result and an actual demand, and deploying the optimized LSTM model to an actual system for fault diagnosis. According to the invention, the fault diagnosis of the aviation equipment can be more accurate and rapid.

Description

Technical Field

[0001] The invention relates to the field of fault diagnosis, and in particular to a fault diagnosis method based on a long short-term memory neural network. Background Art

[0002] Long short-term memory neural network, LSTM for short, is a special recurrent neural network (RNN) designed to solve the gradient vanishing and gradient exploding problems encountered by traditional RNN when processing long sequence data. LSTM manages the flow of information by introducing internal states and three special "gate" structures, so that it can capture long-term dependencies in the sequence; Long short-term memory neural network (LSTM) has important applications in aviation equipment fault diagnosis. As a special recurrent neural network (RNN), LSTM is good at processing time series data and capturing long-term dependencies, which makes it very suitable for fault diagnosis of aviation equipment with complex time series characteristics. In the process of using long short-term memory neural network for aviation equipment fault diagnosis, fault diagnosis method will be used.

[0003] The existing fault diagnosis method uses a long short-term memory neural network with low reliability, which leads to large deviation in the final diagnosis result and has a certain impact on the use of the fault diagnosis method. Therefore, a fault diagnosis method based on a long short-term memory neural network is proposed. Summary of the invention

[0004] The technical problem to be solved by the present invention is: how to solve the problem that the reliability of the long short-term memory neural network established in the existing fault diagnosis method is low, resulting in a large deviation in the final diagnosis result, which has a certain impact on the use of the fault diagnosis method. A fault diagnosis method based on the long short-term memory neural network is provided.

[0005] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps: Step 1: Collect aviation equipment related information and perform data preprocessing. Data preprocessing includes data collection, data cleaning and data normalization; Step 2: Feature extraction, feature extraction of preprocessed data; Step 3: Model construction: construct an LSTM model based on the extracted features; Step 4: Model training: training the constructed LSTM model; Step 5: Fault diagnosis, model testing, and obtaining model test results; Step 6: Optimize the model. According to the test results and actual needs, optimize the LSTM model and deploy the optimized LSTM model to the actual system for fault diagnosis.

[0006] Furthermore, the specific process of the data preprocessing is as follows: First, data collection is performed by setting various sensors in the aviation equipment to be monitored to collect sensor data, obtain sensor data, and collect the fault records of the aviation equipment in the past; Afterwards, the collected sensor data and fault records are cleaned to obtain the cleaning completion data; After normalizing the cleaned data, the data preprocessing process is completed.

[0007] Furthermore, the specific process of the feature extraction is as follows: extracting time series features from the preprocessed data, and extracting required features from the time series data using the sequence learning ability of the LSTM network; And introduce the attention mechanism to enhance the LSTM network's attention to the required features.

[0008] Furthermore, the specific process of introducing the attention mechanism is as follows: Design an attention layer and add an attention layer after the output layer of the LSTM network; The attention layer calculates the importance weight of each time step; During training, the model learns how to dynamically adjust the importance weights based on the input data; Through the back-propagation algorithm, the model continuously optimizes the parameters of the attention layer to increase the focus on the required features; After training is completed, the desired features are extracted by analyzing the weights of the attention layer.

[0009] Furthermore, the specific process of constructing the model is as follows: First, construct the LSTM network layer and design the hierarchical structure of the LSTM network, including the input layer, hidden layer, and output layer; The LSTM unit in the hidden layer has a forgetting function, which is used to forget useless information when necessary; Then select the optimization algorithm to optimize the model. After the optimization is completed, export the model to obtain the LSTM model.

[0010] Furthermore, the forgetting function of LSTM is evaluated, and the next step is allowed after the evaluation is passed. The specific process of evaluating the forgetting function of LSTM is as follows: at least m groups of evaluation data for evaluating the forgetting function are preset; Use m groups of evaluation data to test the forgetting function of LSTM, and record the test results of each group of evaluation data when LSTM performs the forgetting function test. The test results include the accuracy of useless information forgetting and the time taken to complete the useless information forgetting; When the accuracy of forgetting useless information in m test results is greater than the preset value, m and The number of useless information forgotten for a period of time less than the preset time is m and m, it means retesting; When the number of useless information forgetting accuracy greater than the preset value among m test results is less than m or the time taken to forget useless information is less than the preset time. The number is less than When m, it means the test fails; The specific process of selecting the optimization algorithm is as follows: The optimization algorithms used include gradient descent algorithm, stochastic gradient descent algorithm, momentum algorithm and Adam algorithm; Use simulated data to perform simulation optimization evaluation on the gradient descent algorithm, stochastic gradient descent algorithm, momentum algorithm and Adam algorithm, and obtain the optimization range of the gradient descent algorithm, the optimization range of the stochastic gradient descent algorithm, the optimization range of the momentum algorithm and the optimization range of the Adam algorithm; Then, the algorithm with the largest optimization range among the optimization range of the gradient descent algorithm, the optimization range of the stochastic gradient descent algorithm, the optimization range of the momentum algorithm and the optimization range of the Adam algorithm is selected as the final optimization algorithm.

[0011] Furthermore, the specific process of the model training is as follows: The data set is set, and the training set is extracted from the data set; Then, the training set data is used to train the LSTM network, and the parameters of the network are updated through an iterative optimization algorithm; During training, the performance of the model on the validation set is regularly evaluated to prevent overfitting.

[0012] Furthermore, the specific process of fault diagnosis is as follows: Set up a data set, extract a test set from the data set, perform fault diagnosis tests on the trained LSTM model on the test set, and obtain test results; Determine the type and cause of the fault based on the output test results of the model; Based on the diagnosis results, take appropriate measures to repair or prevent the occurrence of faults.

[0013] The specific process of fault diagnosis testing on the trained LSTM model on the test set is as follows; First, select the evaluation indicators, which include accuracy, precision and recall; According to the selected evaluation indicators, calculate the indicators of the model on the test set, and calculate the accuracy, precision and recall; Afterwards, the accuracy, precision and recall are imported into the preset fault database to match the fault content corresponding to the accuracy, precision and recall.

[0014] Furthermore, after the model optimization is completed, subsequent model testing is required. The specific process of subsequent model testing is as follows: At least x groups of test data are set, and the x groups of test data are imported into the optimized model. The optimized model processes the x groups of test data, and the optimized model test data is recorded; The optimized model test data includes diagnostic accuracy and fault diagnosis speed; Then, the single group of optimized model test data is processed to obtain the single group data score; Finally, the average of all single-group data scores is calculated to obtain the comprehensive score. When the comprehensive score is greater than or equal to the preset value, it means that the test has passed and the model will be put into use. When the comprehensive score is less than the preset value, it means that the test fails and the model needs to be optimized.

[0015] Furthermore, the process of obtaining the score of a single group of data is as follows: the fault diagnosis speed is scored, when the diagnosis speed is greater than the preset value a1, its score is f1, when the diagnosis speed is between the preset values ​​a1 and a2, its score is f2, when the diagnosis speed is less than the preset value a2, its score is f3; The diagnostic accuracy is scored. When the diagnostic accuracy is greater than the preset value b1, the score is y1; when the diagnostic accuracy is between the preset values ​​b1 and b2, the score is y2; when the diagnostic accuracy is less than the preset value b2, the score is y3; a1>a2, f1<f2<f3, b1>b2, y1>y2>y3; The fault diagnosis speed is marked as P1, the diagnosis accuracy is marked as P2, and a correction value G1 is assigned to P1 and a correction value G2 is assigned to P2, G1+G2=1, G2>G1; Through the formula G1*P1+G2*P2=Gg, we can get the score of a single group of data.

[0016] Compared with the prior art, the present invention has the following advantages: the fault diagnosis method based on long short-term memory neural network can ensure the accuracy and consistency of data, reduce the influence of noise and outliers on model training, further enhance the LSTM network's ability to capture key features in time series data, and improve the accuracy of fault diagnosis. The construction of the LSTM network layer and the selection of the optimization algorithm can be flexibly adjusted according to the needs of the specific application scenario to achieve the best diagnostic effect. The evaluation of the forgetting function ensures that the LSTM network can effectively forget useless information and retain key information useful for fault diagnosis. The simulation evaluation and selection of the optimization algorithm enable the model to converge faster during the training process and improve the training efficiency. Regularly evaluating the performance of the model on the validation set effectively prevents the occurrence of overfitting and improves the generalization ability of the model. The accuracy and timeliness of fault diagnosis are better. The fault diagnosis test of the trained LSTM model on the test set can accurately determine the type and cause of the fault, providing strong support for subsequent repair or preventive measures. The selection and calculation of the evaluation index make the fault diagnosis result more objective and accurate. By deploying the optimized LSTM model to the actual system for fault diagnosis, the automation and intelligence of fault prediction and diagnosis are achieved, which improves work efficiency and accuracy, making the system more worthy of promotion and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall structural diagram of the present invention. DETAILED DESCRIPTION

[0018] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0019] like Figure 1 As shown, this embodiment provides a technical solution: a fault diagnosis method based on a long short-term memory neural network, comprising the following steps: Step 1: Collect aviation equipment related information and perform data preprocessing. Data preprocessing includes data collection, data cleaning and data normalization; Step 2: Feature extraction, feature extraction of preprocessed data; Step 3: Model construction: construct an LSTM model based on the extracted features; Step 4: Model training: training the constructed LSTM model; Step 5: Fault diagnosis, model testing, and obtaining model test results; Step 6: Optimize the model. According to the test results and actual needs, optimize the LSTM model and deploy the optimized LSTM model to the actual system for fault diagnosis.

[0020] Furthermore, the specific process of the data preprocessing is as follows: First, data collection is performed by setting various sensors in the aviation equipment to be monitored to collect sensor data, obtain sensor data, and collect the fault records of the aviation equipment in the past; Afterwards, the collected sensor data and fault records are cleaned to obtain the cleaning completion data; After normalizing the cleaned data, the data preprocessing process is completed; The specific contents of data cleaning include: removing noise, using filtering technology or signal processing technology to remove noise in the data, and methods such as sliding average and Kalman filtering can be used.

[0021] Handle outliers. Use statistical methods (such as the 3σ principle) or machine learning methods (such as isolation forests) to detect and handle outliers. Outliers may be caused by sensor failures or data recording errors.

[0022] Handling missing values, identifying missing values ​​in the data, you can use interpolation methods (such as linear interpolation, polynomial interpolation) or regression methods to fill the missing values.

[0023] For data with too many missing values, you may need to consider deleting or re-collecting.

[0024] Data normalization: Select a normalization method. Commonly used normalization methods include Min-Max Scaling, Z-score normalization, etc.

[0025] Min-max scaling scales the data to between 0 and 1 and is suitable for situations where the data distribution range is known.

[0026] Z-score standardization converts data into a distribution with a mean of 0 and a standard deviation of 1. It is suitable for situations where the data distribution is unknown or varies greatly.

[0027] Implement normalization: Normalize each sensor data column separately.

[0028] Make sure the normalized data is consistent before model training.

[0029] The specific process of feature extraction is as follows: extract time series features from the preprocessed data, and use the sequence learning ability of the LSTM network to extract the required features from the time series data; And introduce the attention mechanism to enhance the LSTM network's attention to the required features; Through the above process, dynamic changes can be captured. Time series feature extraction can capture dynamic changes in time series data, which is crucial for fault diagnosis. By extracting time series features, we can better understand the changes in data over time and thus identify fault modes more accurately.

[0030] Improving diagnostic accuracy, extracting time series features helps the LSTM network learn the underlying laws and patterns in the data more accurately. These features contain key information in time series data and are of great significance for improving the accuracy of fault diagnosis.

[0031] The LSTM network can effectively capture long-term dependencies in time series data through its special structure (such as input gate, forget gate and output gate) and cell state. This is particularly useful for identifying failure modes that span longer time periods.

[0032] Traditional recurrent neural networks (RNNs) are prone to gradient vanishing problems when processing long sequences, making them difficult to train. The LSTM network effectively alleviates this problem by introducing a gating mechanism, allowing the network to process longer sequence data.

[0033] The attention mechanism can guide the LSTM network to pay more attention to the key features in the time series data. By calculating the importance weight of each time step, the attention layer enables the network to dynamically adjust the attention to the input data, thereby more effectively extracting and utilizing key information.

[0034] After introducing the attention mechanism, we analyze the weights of the attention layer to understand which features the model focuses on when making decisions. This helps us better understand the decision-making process of the model and improve the interpretability of the model.

[0035] The attention mechanism can enhance the sensitivity of the LSTM network to key features, thereby improving the performance of fault diagnosis. By more accurately capturing and utilizing key information, the model can more accurately identify fault types and predict fault development trends.

[0036] The feature extraction process combines the advantages of time series feature extraction, LSTM network's sequence learning ability and attention mechanism, which can more effectively extract key features from time series data and improve the accuracy and performance of fault diagnosis. This is of great significance for fault diagnosis tasks in practical applications and helps to realize the automation and intelligence of fault prediction and diagnosis.

[0037] The specific process of introducing the attention mechanism is as follows: Design an attention layer and add an attention layer after the output layer of the LSTM network; The attention layer calculates the importance weight of each time step; During training, the model learns how to dynamically adjust the importance weights based on the input data; Through the back-propagation algorithm, the model continuously optimizes the parameters of the attention layer to increase the focus on the required features; After training is completed, the required features are extracted by analyzing the weights of the attention layer; The above process can improve model performance and enhance feature attention. The attention mechanism can guide the model to pay more attention to the key features in the input data by calculating the importance weight of each time step. This helps the model learn and use this information more effectively during the training process, thereby improving the performance of the model.

[0038] Optimize resource allocation. The attention mechanism allows the model to dynamically allocate computing resources according to the needs of the current task when processing input data. This helps the model to be more efficient when processing complex tasks and reduce unnecessary computing overhead.

[0039] Improve the interpretability of the model. The weights of the attention layer can intuitively reflect the model's attention to different parts when processing input data. By visualizing these weights, we can better understand the model's decision-making process and improve the interpretability of the model.

[0040] Better auxiliary fault diagnosis, in applications such as fault diagnosis, by analyzing the weights of the attention layer, we can understand which features the model focuses on when identifying fault modes. This helps us locate the source of the fault more accurately and improve the accuracy and efficiency of fault diagnosis.

[0041] Enhance the model's generalization ability to adapt to different input lengths and reduce the risk of overfitting; The attention mechanism does not rely on a fixed length of input data, which enables the model to handle input sequences of different lengths. This helps the model be more flexible and robust in dealing with different scenarios and tasks.

[0042] The attention mechanism helps reduce the model's reliance on noise and irrelevant information by guiding the model to focus on key features in the input data, thereby reducing the risk of overfitting.

[0043] The attention mechanism helps the model learn the underlying rules and patterns in the input data more quickly by dynamically adjusting the model's attention to the input data. This can speed up the model training process and improve training efficiency.

[0044] When processing long sequences, traditional recurrent neural networks (RNNs) are prone to the vanishing gradient problem. The attention mechanism helps alleviate this problem by guiding the model to focus on key features, allowing the model to better process long sequence data.

[0045] The specific process of introducing the attention mechanism brings many benefits, including improving model performance, improving model interpretability, enhancing model generalization ability, and optimizing the training process. These benefits make the model more efficient, accurate, and robust when dealing with complex tasks.

[0046] The specific process of model construction is as follows: First, construct the LSTM network layer and design the hierarchical structure of the LSTM network, including the input layer, hidden layer, and output layer; The LSTM unit in the hidden layer has a forgetting function, which is used to forget useless information when necessary; Then select the optimization algorithm to optimize the model. After the optimization is completed, export the model and obtain the LSTM model. The above process can reduce information redundancy. The forgetting function allows the LSTM unit to forget useless or redundant information when needed, which helps to reduce the computational burden of the model and improve the operating efficiency of the model.

[0047] The robustness of the model is enhanced. By forgetting useless information, the LSTM network can better cope with noise and outliers in the input data, thereby improving the robustness and stability of the model.

[0048] The forgetting function helps the LSTM network focus more on learning the key features in the input data, which is crucial to improving the prediction accuracy and generalization ability of the model; It can speed up model training. Appropriate optimization algorithms can speed up the model training process and reduce training time. This is especially important for large-scale data sets and complex models.

[0049] Improve training efficiency. The optimization algorithm adjusts the model parameters to minimize the loss function, thereby improving the training efficiency of the model. This helps the model learn the potential rules and patterns in the input data more quickly.

[0050] The forgetting function of LSTM is evaluated. After passing the evaluation, it is allowed to proceed to the next step. The specific process of evaluating the forgetting function of LSTM is as follows: at least m groups of evaluation data for forgetting function evaluation are preset; Use m groups of evaluation data to test the forgetting function of LSTM, and record the test results of each group of evaluation data when LSTM performs the forgetting function test. The test results include the accuracy of useless information forgetting and the time taken to complete the useless information forgetting; When the accuracy of forgetting useless information in m test results is greater than the preset value, m and The number of times that the useless information is forgotten is less than the preset time. m and m, it means retesting; When the number of useless information forgetting accuracy greater than the preset value among m test results is less than m or the time taken to forget useless information is less than the preset time. The number is less than When m, it means the test fails; The specific process of selecting the optimization algorithm is as follows: The optimization algorithms used include gradient descent algorithm, stochastic gradient descent algorithm, momentum algorithm and Adam algorithm; Use simulated data to perform simulation optimization evaluation on the gradient descent algorithm, stochastic gradient descent algorithm, momentum algorithm and Adam algorithm, and obtain the optimization range of the gradient descent algorithm, the optimization range of the stochastic gradient descent algorithm, the optimization range of the momentum algorithm and the optimization range of the Adam algorithm; Then, the algorithm with the largest optimization range among the optimization range of the gradient descent algorithm, the optimization range of the stochastic gradient descent algorithm, the optimization range of the momentum algorithm and the optimization range of the Adam algorithm is selected as the final optimization algorithm; By presetting multiple sets of evaluation data and conducting tests, the ability of the LSTM network to forget useless information can be comprehensively evaluated. This helps ensure that the LSTM network can effectively filter out irrelevant information in practical applications and improve the accuracy and efficiency of the model; The test results include the accuracy of useless information forgetting and the time it takes to complete useless information forgetting. These two indicators can intuitively reflect the performance of the LSTM network in the forgetting function. By comparing the preset values ​​and the actual test results, it can be clearly judged whether the forgetting function of the LSTM network meets the standard; The evaluation process is strict and comprehensive. When the test results do not meet the requirements, the test will be retested or judged as failed. This helps to timely discover and solve the problems of LSTM network in forgetting function and improve the reliability and stability of the model.

[0051] By comparing the optimization ranges of different optimization algorithms, we can select the optimization algorithm that best suits the LSTM network. This helps improve the training efficiency and prediction accuracy of the model, making the model perform better in practical applications; Different optimization algorithms have different computational complexity and resource consumption. By selecting the algorithm with the largest optimization range, you can save as much computing resources and time as possible while ensuring model performance.

[0052] The specific process of model training is as follows: The data set is set, and the training set is extracted from the data set; Then, the training set data is used to train the LSTM network, and the parameters of the network are updated through an iterative optimization algorithm; During training, the performance of the model on the validation set is regularly evaluated to prevent overfitting.

[0053] The specific process of fault diagnosis is as follows: Set up a data set, extract a test set from the data set, perform fault diagnosis tests on the trained LSTM model on the test set, and obtain test results; Determine the type and cause of the fault based on the output test results of the model; Based on the diagnosis results, take appropriate measures to repair or prevent the occurrence of faults.

[0054] The specific process of fault diagnosis testing on the trained LSTM model on the test set is as follows; First, select the evaluation indicators, which include accuracy, precision and recall; According to the selected evaluation indicators, calculate the indicators of the model on the test set, and calculate the accuracy, precision and recall; Then, the accuracy, precision and recall are imported into the preset fault database to match the fault content corresponding to the accuracy, precision and recall; LSTM (Long Short-Term Memory) has the ability to capture long-term dependencies in time series data, which gives it a unique advantage in the field of fault diagnosis. By performing fault diagnosis tests with a trained LSTM model, the type and cause of the fault can be identified more accurately.

[0055] Accuracy, precision, and recall are important indicators for measuring model performance. In fault diagnosis, these indicators can fully reflect the model's ability to judge the fault type and the accuracy of identifying the fault cause. By calculating these indicators, the performance of the model can be evaluated more objectively, thereby improving the accuracy of fault diagnosis.

[0056] Importing the calculated accuracy, precision and recall into the preset fault database can quickly match the corresponding fault content. This not only improves the efficiency of fault diagnosis, but also makes the diagnosis results more intuitive and easy to understand.

[0057] The entire fault diagnosis process is automated and intelligent. By testing with a trained LSTM model, the test results can be automatically output, and the fault type and cause can be determined based on the evaluation indicators. This avoids the steps of manual analysis and judgment required in traditional fault diagnosis methods, and improves the efficiency and accuracy of fault diagnosis.

[0058] Reduce human interference. Traditional fault diagnosis methods are easily interfered by human factors, such as lack of experience and misjudgment. Using the LSTM model for fault diagnosis can reduce the interference of human factors and make the diagnosis results more objective and reliable.

[0059] It is convenient to record and track faults. Through the preset fault database, the fault content and diagnosis results can be easily recorded and tracked. This helps enterprises to establish a complete fault management system and provide a basis for subsequent fault analysis and improvement.

[0060] After the model optimization is completed, subsequent model testing is required. The specific process of subsequent model testing is as follows: At least x groups of test data are set, and x groups of test data are imported into the optimized model. The optimized model processes x groups of test data, and the optimized model test data is recorded, x ≥ 10; The optimized model test data includes diagnostic accuracy and fault diagnosis speed; Then, the single group of optimized model test data is processed to obtain the single group data score; Finally, the average of all single-group data scores is calculated to obtain the comprehensive score. When the comprehensive score is greater than or equal to the preset value, it means that the test has passed and the model will be put into use. When the comprehensive score is less than the preset value, it means that the test fails and the model needs to be optimized.

[0061] The process of obtaining the score of a single group of data is as follows: score the fault diagnosis speed. When the diagnosis speed is greater than the preset value a1, its score is f1; when the diagnosis speed is between the preset values ​​a1 and a2, its score is f2; when the diagnosis speed is less than the preset value a2, its score is f3; The diagnostic accuracy is scored. When the diagnostic accuracy is greater than the preset value b1, the score is y1; when the diagnostic accuracy is between the preset values ​​b1 and b2, the score is y2; when the diagnostic accuracy is less than the preset value b2, the score is y3; a1>a2, f1<f2<f3, b1>b2, y1>y2>y3; The fault diagnosis speed is marked as P1, the diagnosis accuracy is marked as P2, and a correction value G1 is assigned to P1 and a correction value G2 is assigned to P2, G1+G2=1, G2>G1; Through the formula G1*P1+G2*P2=Gg, the single group data score Gg is obtained; The above process enables a comprehensive evaluation of the model performance. Setting at least x sets of test data can fully cover various situations that the model may encounter, thereby more accurately evaluating the model's performance. This helps to discover the performance differences of the model in different scenarios and provide direction for subsequent optimization.

[0062] Comprehensive scoring mechanism: By calculating the average of all single-group data scores to obtain a comprehensive score, the overall performance of the model can be more objectively reflected. This comprehensive scoring mechanism not only takes into account the diagnostic accuracy, but also the fault diagnosis speed, making the evaluation results more comprehensive and accurate.

[0063] Segmented scoring: Segmented scoring of fault diagnosis speed and diagnostic accuracy can more carefully reflect the performance of the model at different performance levels. This helps to find out where the model is deficient and provide specific improvement directions for subsequent optimization.

[0064] By assigning different correction values ​​G1 and G2 to the fault diagnosis speed and diagnostic accuracy, and calculating the score of a single set of data according to the formula G1P1+G2P2=Gg, the score weight can be adjusted more flexibly, making the score result more in line with actual needs. This correction value mechanism helps to balance the weights between different performance indicators, making the evaluation results more reasonable.

[0065] Improve model reliability and stability, and strictly enforce testing standards: By setting preset values ​​and comprehensive score thresholds, you can ensure that the model meets certain performance standards during the test process. Only when the comprehensive score is greater than or equal to the preset value can the model be put into use, which helps improve the reliability and stability of the model.

[0066] Continuous optimization: When the comprehensive score is less than the preset value, it means that the test fails and the model needs to be optimized. This continuous optimization mechanism ensures that the model is continuously improved and perfected in practical applications, thereby improving its adaptability and robustness.

[0067] The collection covers various sensor data and historical fault records in the aviation equipment that needs to be monitored, providing a rich information basis for subsequent fault diagnosis.

[0068] Data cleaning and normalization ensure the accuracy and consistency of the data and reduce the impact of noise and outliers on model training.

[0069] The introduction of time series feature extraction and attention mechanism further enhances the LSTM network's ability to capture key features in time series data and improves the accuracy of fault diagnosis.

[0070] The construction of the LSTM network layer and the selection of the optimization algorithm can be flexibly adjusted according to the needs of specific application scenarios to achieve the best diagnostic effect.

[0071] The evaluation of the forgetting function ensures that the LSTM network can effectively forget useless information and retain key information that is useful for fault diagnosis.

[0072] The simulation evaluation and selection of optimization algorithms enable the model to converge faster during training and improve training efficiency.

[0073] The partitioning of the data set and the use of iterative optimization algorithms ensure that the model can learn the underlying laws and patterns of the data during the training process.

[0074] Regularly evaluating the performance of the model on the validation set effectively prevents overfitting and improves the generalization ability of the model.

[0075] Performing fault diagnosis tests on the trained LSTM model on the test set can accurately determine the type and cause of the fault, providing strong support for subsequent repair or preventive measures.

[0076] The selection and calculation of evaluation indicators make the fault diagnosis results more objective and accurate.

[0077] Subsequent model testing and optimization processes ensure that the model can continue to maintain high performance in practical applications.

[0078] The calculation of the comprehensive score and the scoring of the fault diagnosis speed and accuracy provide a clear direction and basis for further optimization of the model.

[0079] Strong practical applicability: This method is not only suitable for fault diagnosis of aviation equipment, but can also be widely used in fault diagnosis in other fields and has wide applicability.

[0080] By deploying the optimized LSTM model to the actual system for fault diagnosis, the automation and intelligence of fault prediction and diagnosis are achieved, and the work efficiency and accuracy are improved.

[0081] In summary, the above LSTM-based fault diagnosis method realizes efficient and accurate diagnosis of aviation equipment faults through a series of scientific and systematic steps, providing strong technical support for the operation, maintenance and safety management of equipment.

[0082] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0083] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0084] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A fault diagnosis method based on long short-term memory neural network, characterized in that: The following steps are involved: Step 1: Collect aviation equipment related information and perform data preprocessing; Step 2: Feature extraction, feature extraction of preprocessed data; Step 3: Model construction: construct an LSTM model based on the extracted features; Step 4: Model training: training the constructed LSTM model; Step 5: Fault diagnosis, model testing, and obtaining model test results; Step 6: Optimize the model. According to the test results and actual needs, optimize the LSTM model and deploy the optimized LSTM model to the actual system for fault diagnosis.

2. A fault diagnosis method based on long short-term memory neural network according to claim 1, characterized in that: The specific process of data preprocessing is as follows: First, data collection is performed by setting various sensors in the aviation equipment to be monitored to collect sensor data, obtain sensor data, and collect the fault records of the aviation equipment in the past; Afterwards, the collected sensor data and fault records are cleaned to obtain the cleaning completion data; After normalizing the cleaned data, the data preprocessing process is completed.

3. The fault diagnosis method based on long short-term memory neural network according to claim 1 is characterized in that: The specific process of feature extraction is as follows: extract time series features from the preprocessed data, and use the sequence learning ability of the LSTM network to extract the required features from the time series data; And introduce the attention mechanism to enhance the LSTM network's attention to the required features.

4. A fault diagnosis method based on long short-term memory neural network according to claim 3, characterized in that: The specific process of introducing the attention mechanism is as follows: Design an attention layer and add an attention layer after the output layer of the LSTM network; The attention layer calculates the importance weight of each time step; During training, the model learns how to dynamically adjust the importance weights based on the input data; Through the back-propagation algorithm, the model continuously optimizes the parameters of the attention layer to increase the focus on the required features; After training is completed, the desired features are extracted by analyzing the weights of the attention layer.

5. The fault diagnosis method based on long short-term memory neural network according to claim 1 is characterized in that: The specific process of model construction is as follows: First, construct the LSTM network layer and design the hierarchical structure of the LSTM network, including the input layer, hidden layer, and output layer; The LSTM unit in the hidden layer has a forgetting function, which is used to forget useless information when necessary; Then select the optimization algorithm to optimize the model. After the optimization is completed, export the model to obtain the LSTM model.

6. A fault diagnosis method based on long short-term memory neural network according to claim 5, characterized in that: The forgetting function of LSTM is evaluated. After passing the evaluation, it is allowed to proceed to the next step. The specific process of evaluating the forgetting function of LSTM is as follows: at least m groups of evaluation data for forgetting function evaluation are preset; Use m groups of evaluation data to test the forgetting function of LSTM, and record the test results of each group of evaluation data when LSTM performs the forgetting function test. The test results include the accuracy of useless information forgetting and the time taken to complete the useless information forgetting; Extract m test results. When the accuracy of useless information forgetting in the m test results is greater than the preset value, and the time taken to forget useless information is less than the preset time, it means that the test has passed. When the accuracy of forgetting useless information in m test results is greater than the preset value, m and The number of useless information forgotten for a period of time less than the preset time is m and m, it means retesting; When the number of useless information forgetting accuracy greater than the preset value among m test results is less than m or the time taken to forget useless information is less than the preset time. The number is less than When m, it means the test fails; The specific process of selecting the optimization algorithm is as follows: The optimization algorithms used include gradient descent algorithm, stochastic gradient descent algorithm, momentum algorithm and Adam algorithm; Use simulated data to perform simulation optimization evaluation on the gradient descent algorithm, stochastic gradient descent algorithm, momentum algorithm and Adam algorithm, and obtain the optimization range of the gradient descent algorithm, the optimization range of the stochastic gradient descent algorithm, the optimization range of the momentum algorithm and the optimization range of the Adam algorithm; Then, the algorithm with the largest optimization range among the optimization range of the gradient descent algorithm, the optimization range of the stochastic gradient descent algorithm, the optimization range of the momentum algorithm and the optimization range of the Adam algorithm is selected as the final optimization algorithm.

7. The fault diagnosis method based on long short-term memory neural network according to claim 1 is characterized in that: The specific process of model training is as follows: The data set is set, and the training set is extracted from the data set; Then, the training set data is used to train the LSTM network, and the parameters of the network are updated through an iterative optimization algorithm; During training, the performance of the model on the validation set is regularly evaluated to prevent overfitting.

8. A fault diagnosis method based on long short-term memory neural network according to claim 7, characterized in that: The specific process of fault diagnosis is as follows: Set up a data set, extract a test set from the data set, perform fault diagnosis tests on the trained LSTM model on the test set, and obtain test results; Determine the type and cause of the fault based on the output test results of the model; According to the diagnosis results, take appropriate measures to repair or prevent the occurrence of faults; The specific process of fault diagnosis testing on the trained LSTM model on the test set is as follows; First, select the evaluation indicators, which include accuracy, precision and recall; According to the selected evaluation indicators, calculate the indicators of the model on the test set, and calculate the accuracy, precision and recall; Afterwards, the accuracy, precision and recall are imported into the preset fault database to match the fault content corresponding to the accuracy, precision and recall.

9. The fault diagnosis method based on long short-term memory neural network according to claim 1 is characterized in that: After the model optimization is completed, subsequent model testing is required. The specific process of subsequent model testing is as follows: At least x groups of test data are set, and the x groups of test data are imported into the optimized model. The optimized model processes the x groups of test data, and the optimized model test data is recorded; The optimized model test data includes diagnostic accuracy and fault diagnosis speed; Then, the single group of optimized model test data is processed to obtain the single group data score; Finally, the average of all single-group data scores is calculated to obtain the comprehensive score. When the comprehensive score is greater than or equal to the preset value, it means that the test has passed and the model will be put into use. When the comprehensive score is less than the preset value, it means that the test fails and the model needs to be optimized.

10. A fault diagnosis method based on long short-term memory neural network according to claim 9, characterized in that: The process of obtaining the score of a single group of data is as follows: score the fault diagnosis speed. When the diagnosis speed is greater than the preset value a1, its score is f1; when the diagnosis speed is between the preset values ​​a1 and a2, its score is f2; when the diagnosis speed is less than the preset value a2, its score is f3; The diagnostic accuracy is scored. When the diagnostic accuracy is greater than the preset value b1, the score is y1; when the diagnostic accuracy is between the preset values ​​b1 and b2, the score is y2; when the diagnostic accuracy is less than the preset value b2, the score is y3; a1>a2, f1<f2<f3, b1>b2, y1>y2>y3; The fault diagnosis speed is marked as P1, the diagnosis accuracy is marked as P2, and a correction value G1 is assigned to P1 and a correction value G2 is assigned to P2, G1+G2=1, G2>G1; Through the formula G1*P1+G2*P2=Gg, we can get the score of a single group of data.

Images