Civil aircraft frame structure crack repair decision-making method fusing physical information and case data

Abstract

The application discloses a civil aircraft partition frame structure crack repair decision-making method fusing physical information and case data, and solves the problem that the structure repair manual is not mature enough and repair cases are insufficient at the initial stage of civil aircraft operation, so that the partition frame structure crack repair is difficult to be decided. Through learning the current mainstream aircraft partition frame structure crack repair cases from manufacturers, operators and maintenance units, a civil aircraft partition frame structure crack repair historical case library is constructed. Test pieces made of the same material as the partition frame structure of the civil aircraft are selected, and equal-amplitude fatigue crack propagation experiments are carried out according to the actual operation load spectrum, and the crack size and corresponding load cycle number are recorded in detail. A PINN deep learning model is selected, Paris law is added as a model constraint, physical information is embedded into the network through a compound loss function, and crack residual life prediction is realized. The pre-training models of PINN, BERT and ViT are used to analyze the structured data, text data and image data, and on this basis, the twin architecture is used to analyze the multi-modal and multi-dimensional similarity between cases, so that a similar case repair scheme is matched for a new partition frame structure crack damage, so that the civil aircraft partition frame structure crack repair decision-making based on cases is realized. The application realizes accurate and reliable repair case matching, supports the civil aircraft partition frame structure repair decision-making, and improves the decision-making efficiency and commercial competitiveness.

Description

Technical Field

[0001] This invention relates to the field of civil aircraft operation support technology, specifically to a decision-making method for repairing cracks in the frame structure of a civil aircraft that integrates physical information and case data. Background Technology

[0002] Currently, the repair of bulkhead structures in civil aircraft primarily relies on structural repair manuals. However, these manuals typically have a long maturation period after newly developed aircraft enter service. Therefore, during the initial operational phase, numerous bulkhead crack repair tasks exceeding the scope of the manuals arise, requiring engineers to comprehensively consider aircraft design and material properties, damage assessment, aircraft operational records, operating environment, maintenance resources, and capabilities to develop repair plans. Developing repair plans for these tasks is challenging, highly dependent on experience, and involves lengthy approval cycles, severely impacting repair decision-making efficiency and resulting in low operational availability for civil aircraft. Therefore, there is an urgent need to develop a method that can comprehensively utilize domain knowledge and current mainstream aircraft bulkhead crack repair case studies to achieve rapid response and efficient decision-making in civil aircraft bulkhead crack repair.

[0003] Deep learning, with its excellent knowledge embedding and data mining capabilities, has become an important research tool in engineering and technology in recent years. This makes it possible to make decisions on the repair of structural cracks in civil aircraft using physical information and case data. The physical information neural network model is a deep learning model that imposes physical information constraints during model training. It can accurately predict the remaining life of cracks in bulkhead structures by embedding fracture mechanics physical information based on fully utilizing structured case data. Simultaneously, pre-trained models using text and images can fully extract features from case text and image data even with limited sample sizes. Through the processing and analysis of physical information and case data, repair schemes can be matched for bulkhead structure crack repair tasks, supporting decision-making on civil aircraft bulkhead structure crack repair, improving decision-making efficiency, and enhancing the commercial competitiveness of civil aircraft. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for repairing cracks in the frame structure of civil aircraft that integrates physical information and case data, thereby shortening the repair plan formulation cycle and improving the efficiency of repair decision-making.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for decision-making regarding the repair of cracks in the bulkhead structure of a civil aircraft, which integrates physical information and case data, is characterized by comprising the following steps:

[0007] Step 1: Integrate and learn from current mainstream aircraft frame structure crack repair cases to build a historical case library of civil aircraft frame structure crack repair.

[0008] Step 2: Based on the civil aircraft frame structure, select crack tensile test specimens of the same material, simulate the fatigue load spectrum in actual operation, conduct constant amplitude fatigue crack propagation experiments, and record the crack length and corresponding load cycle number in detail.

[0009] Step 3: Select the PINN deep learning model, input the experimental data recorded in Step 2 and the structured case data collected in Step 1 into the model, add Paris's law as a model constraint, and embed physical information into the network through a composite loss function to achieve crack remaining life prediction.

[0010] Step 4: Select a pre-trained BERT model, input the case text data collected in Step 1 into the model for fine-tuning, and extract the features of the case text data.

[0011] Step 5: Select the pre-trained ViT model, input the case image data collected in Step 1 into the model for fine-tuning, and extract the features of the case image data.

[0012] Step 6: Based on the analysis results of Steps 3, 4, and 5, use a twin architecture to perform multimodal and multidimensional similarity analysis between cases, match similar case repair schemes for new bulkhead structural crack damage, and make case-based decisions on the repair of civil aircraft bulkhead structural cracks.

[0013] Step 1 of this invention integrates and learns repair case data of cracks in the bulkhead structure of current mainstream aircraft models from manufacturers, operators and maintenance units, and uses SVM to perform independence and correlation analysis between indicators, retains representative indicators, removes redundant data, and constructs a civil aircraft bulkhead structure crack repair case library containing structured data, text data and image data.

[0014] The data learned from the manufacturer is mainly structured data, which consists of various initial data of civil aircraft determined by design and manufacturing, including: design and manufacturing data, geometric and dimensional data, flight performance data, structural health monitoring data, material property data, damage tolerance and fatigue analysis data, and technical publications.

[0015] The data learned from the operator includes structured data, image data, and text data, which are data generated by the aircraft during operation as recorded by the operator, including: aircraft operation data, aircraft repair records, specific damage information, damage location information, damage discovery records, and damage location images;

[0016] The data learned from the maintenance unit is mainly text data, which involves specific repair information on crack damage in the frame structure, including: description of repair methods, specific repair plans, repair materials, adhesive bonding process, surface treatment and protective measures, and inspection requirements after repair.

[0017] In step 2 of this invention, a test piece made of the same material as the civil aircraft frame structure is selected, and the applied load is determined according to the actual operating environment and frequency. A fatigue crack propagation experiment is conducted, and the crack length and the corresponding load cycle number are recorded.

[0018] Step 3 of this invention involves establishing a fully connected PINN neural network model, using Paris's law of crack propagation in fracture mechanics as a model constraint, embedding physical information into the network through a composite loss function, taking experimental data and structured data from repair cases as inputs, and outputting the remaining life of the crack, establishing the relationship between crack damage and fatigue cycles in the civil aircraft frame structure, and using automatic differentiation to solve the network gradient and adjust the network weights.

[0019] The Paris law for crack propagation is specifically expressed as follows:

[0020]

[0021] In the formula, a is the crack length, n is the number of load cycles, and C and m are material-related parameters;

[0022] The composite loss function is specifically expressed as follows:

[0023] Loss = w Data Loss Data +w Physics Loss Physics

[0024] In the formula, Loss is the loss function of the PINN network. Data For the structured data loss portion of the case, Loss Physics For the physical information loss part, w Data and w Physics These are the weighting coefficients for the structured data loss component and the physical information loss component of the case, respectively.

[0025] In step 4 of this invention, a BERT neural network model is established. The model is initialized using the bert-base-chinese pre-trained weights provided by Google. During the training phase, an MLM training task is constructed using text data from civil aircraft frame structure crack repair cases to fine-tune the model and obtain a language model with characteristics of the civil aircraft frame structure crack repair domain. During the inference phase, the BERT model is used to vectorize the input text and extract the features of the case text data.

[0026] The MLM training task is an unsupervised learning task. The model replaces the segmentation units in the text with a probability of 15%. 80% of the units are replaced with masks, 10% of the units are replaced with any other unit, and the remaining 10% remain unchanged.

[0027] In step 5 of this invention, a ViT neural network model is established. The model is initialized using training weights obtained from training on the image dataset of crack detection in the civil aircraft frame structure. During the inference phase, the ViT model is used to extract multi-level features from the input image and extract features from the case image data.

[0028] In step 6 of this invention, a twin architecture model with two branch networks is established. Each branch network consists of the three neural network models from steps 4, 5, and 6. The two branches are used to perform multimodal and multidimensional processing and analysis on new cases of crack damage to the bulkhead structure and cases in the historical case library, respectively. The outputs of the two branches are then subjected to similarity analysis. Each new case of crack damage to the bulkhead structure is compared with cases in the historical case library to obtain the repair plan for the most similar case, and case-based decision-making for repairing cracks in the bulkhead structure of civil aircraft is carried out.

[0029] This invention addresses a pressing practical engineering problem in the initial operation phase of civil aircraft by disclosing an accurate, reliable, and highly interpretable system solution. It utilizes a deep learning model to embed fracture mechanics knowledge and mine repair case data, integrating physical information and case data to achieve multimodal and multidimensional similarity analysis between repair cases. This enables accurate matching of repair cases, effectively shortens the repair plan development cycle, improves the efficiency of repair decision-making for cracks in civil aircraft frame structures, and effectively supports the commercial operation of civil aircraft. Attached Figure Description

[0030] Figure 1 This invention provides a schematic flowchart of a decision-making method for repairing cracks in the frame structure of a civil aircraft that integrates physical information and case data. Detailed Implementation

[0031] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. However, this description should not be considered a limitation of the present invention, but rather a more detailed description of certain features and technical solutions of the present invention.

[0032] like Figure 1 The diagram shown is a flowchart illustrating a decision-making method for repairing cracks in a civil aircraft bulkhead structure, which integrates physical information and case data, provided by the present invention. The method operates according to the following steps:

[0033] S1. Integrate and learn from repair case data of cracks in the bulkhead structure of current mainstream aircraft models from manufacturers, operators, and maintenance units. Use SVM to analyze the independence and correlation between indicators, retain representative indicators, and remove redundant data to construct a civil aircraft bulkhead structure crack repair case library containing structured data, text data, and image data. Specific steps include:

[0034] S11. The data learned from the manufacturer is mainly structured data, which consists of various initial data of civil aircraft determined by design and manufacturing, including: design and manufacturing data, geometric and dimensional data, flight performance data, structural health monitoring data, material property data, damage tolerance and fatigue analysis data, and technical publications.

[0035] S12. The data learned from the operator includes structured data, image data and text data, which are data generated by the aircraft during operation as recorded by the operator, including: aircraft operation data, aircraft repair records, specific damage information, damage location information, damage discovery records and damage location images;

[0036] S13. The data learned from the maintenance unit is mainly text data, which involves specific repair information on crack damage in the frame structure, including: description of repair methods, specific repair plans, repair materials, adhesive bonding process, surface treatment and protective measures, and inspection requirements after repair.

[0037] S14. Represent a case of repairing a crack in a civil aircraft bulkhead structure using binary tuples, specifically defined as follows:

[0038] Case = {Problem, Solution};

[0039] Case represents a complete case study of a crack repair in the frame structure of a civil aircraft, where C = {C1, C2, C3, ..., C}. i C n} represents the collection of repair cases, specifically the civil aircraft bulkhead structure crack repair case library. Here, n represents the number of cases, and C... i Let i represent the i-th case in the case library C, where i∈{1,2,3,…,n};

[0040] A problem is a description of a repair case represented by a case. i ={S i1 S i2 S i3 S ij S im} represents the set of indicators for the i-th repair case, where m represents the number of indicators for the case, and S ij Let j represent the j-th index in the i-th repair case, where j∈{1,2,3,…,m};

[0041] Solution is the specific repair plan for the repair case represented by Case, where A = {A1, A2, A3, ..., A...} i A n} represents a set of repair solutions. Among them, A i This represents the repair plan for the i-th case in the case library C.

[0042] S2. Based on the civil aircraft frame structure, select tensile test specimens of the same material to simulate the fatigue load spectrum in actual operation. Conduct constant amplitude fatigue crack propagation experiments, recording the crack length and corresponding load cycle number. Record 200 data points for each experiment. The selected test specimens are 50 mm wide and 12 mm thick, and sinusoidal loads are applied with a maximum load of 4.5 kN, a minimum load of 0.9 kN, and a frequency of 20 Hz.

[0043] S3. Establish a fully connected PINN neural network model. The input is experimental data and structured data from repair cases, and the output is the remaining crack life. Set the number of fully connected layers in the neural network to [15, 30, 30, 1], indicating that the PINN network structure has 5 layers, with 15, 30, 30, and 1 nodes in each layer, respectively. Use Paris's law of crack propagation in fracture mechanics as the model constraint, and embed physical information into the network through a composite loss function. Paris's law of crack propagation is specifically expressed as:

[0044]

[0045] In the formula, a is the crack length, n is the number of load cycles, and C and m are material-related parameters;

[0046] The composite loss function is specifically expressed as follows:

[0047] Loss = w Data Loss Data +w Physics Loss Physics ;

[0048] Loss is the loss function of the PINN network. Data For the structured data loss portion of the case, Loss Physics For the physical information loss part, w Data and w Physics These are the weighting coefficients for the structured data loss component and the physical information loss component of the case, respectively.

[0049] S4. Establish a BERT neural network model and extract features from the case text data. Specific steps include:

[0050] S41. Preprocess the text data of the civil aircraft bulkhead structure crack repair case, including removing useless symbols, deleting irrelevant symbols, spaces, numbers, tags, letters and punctuation marks in the text, removing stop words and segmenting them;

[0051] S41. Construct a BERT neural network model with 12 Transformer modules, a hidden layer size of 768, 12 heads in the multi-head attention module, and a maximum input sentence length of 64.

[0052] S42. Initialize the local model parameters using publicly available BERT model weights pre-trained on a large amount of Chinese text. The model weights are the bert-base-chinese pre-trained weights provided by Google.

[0053] S43. Construct an MLM training task using preprocessed text data of civil aircraft frame structure crack repair cases. The MLM training task is an unsupervised learning task. The model replaces the segmentation units in the text with a probability of 15%, 80% of the units are replaced with masks, 10% of the units are replaced with any other unit, and the remaining 10% remain unchanged. Use the constructed MLM training dataset to fine-tune the pre-trained model to obtain a language model with characteristics of the civil aircraft frame structure crack repair domain.

[0054] S44. Use the fine-tuned model for inference. Input the case text into the model, extract the case text data features, and obtain the feature vector sequence.

[0055] S5. Establish a ViT neural network model and extract features from the case image data. Specific steps include:

[0056] S51. Preprocess the image data of the civil aircraft bulkhead structure crack repair case, rotate and scale the image to keep the crack acquisition scale and angle consistent; remove image noise by mean filtering to eliminate obvious noise points; perform local enhancement on the image to highlight the outline of the bulkhead structure crack in the image and improve contrast.

[0057] S52. Establish a ViT neural network model, initialize the model using the training weights obtained from training on the image dataset of crack detection in civil aircraft frame structures, and use the ViT model to extract features from the input image.

[0058] S6. Using twin architecture analysis, multimodal and multidimensional similarity analysis is performed between cases to match repair schemes from similar cases for new bulkhead structural crack damage, thereby achieving case-based decision-making for civil aircraft bulkhead structural crack repair. Specific steps include:

[0059] S61. Establish a Siamese-PINN-BERT-ViT network model with two branch networks. Each branch network is composed of three neural network models from S4, S5, and S6 in parallel. The two branch networks share the model structure and weight parameters.

[0060] S62. Input new cases of crack damage in the diaphragm structure and cases from the historical case library into the two branch networks respectively. Each branch network concatenates the different modal features extracted. The features extracted by the two branch networks are calculated using cosine similarity to obtain the similarity analysis results between cases. The specific similarity calculation method is as follows:

[0061]

[0062] X and Y are the feature vectors output by the two branch networks, respectively;

[0063] S63. Perform similarity analysis on each new case of crack damage in the bulkhead structure and the cases in the historical case database to obtain the repair plan for the most similar case. Refer to the repair plan of the matched case to carry out the repair of the bulkhead structure of the civil aircraft, so as to realize the decision-making of the repair of the crack in the bulkhead structure of the civil aircraft based on physical information and case data.

[0064] This invention discloses a decision-making method for repairing cracks in the frame structure of a civil aircraft that integrates physical information and case data. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several modifications and combinations can be made without departing from the principle of this invention, and these modifications and combinations should also be considered within the scope of protection of this invention.

Claims

1. A decision-making method for repairing cracks in the frame structure of a civil aircraft, integrating physical information and case data, characterized in that... Includes the following steps: Step 1: Integrate and learn from current mainstream aircraft frame structure crack repair cases to build a historical case library of civil aircraft frame structure crack repair. Step 2: Based on the civil aircraft frame structure, select crack tensile test specimens of the same material, simulate the fatigue load spectrum in actual operation, conduct constant amplitude fatigue crack propagation experiments, and record the crack length and corresponding load cycle number in detail. Step 3: Select the PINN deep learning model, input the experimental data recorded in Step 2 and the structured case data collected in Step 1 into the model, add Paris's law as a model constraint, and embed physical information into the network through a composite loss function to achieve crack remaining life prediction. Using experimental data and structured data from repair cases as input, and outputting the remaining life of the crack, the relationship between crack damage and fatigue cycles in the civil aircraft frame structure is established. The network gradient is solved by automatic differentiation, and the network weights are adjusted. Paris's law of crack propagation is specifically expressed as: In the formula, a is the crack length, n is the number of load cycles, and C and m are material-related parameters; The composite loss function is specifically expressed as follows: Loss=w Data Loss Data +w Physics Loss Physics In the formula, Loss is the loss function of the PINN network. Data For the structured data loss portion of the case, Loss Physics For the physical information loss part, w Data and w Physics These are the weighting coefficients for the structured data loss component and the physical information loss component of the case, respectively. Step 4: Select a pre-trained BERT model, input the case text data collected in Step 1 into the model for fine-tuning, and extract the features of the case text data. Step 5: Select a pre-trained ViT model, input the case image data collected in Step 1 into the model for fine-tuning, and extract the features of the case image data. Step 6: Based on the analysis results of Steps 3, 4, and 5, use a twin architecture to perform multimodal and multidimensional similarity analysis between cases, match similar case repair schemes for new bulkhead structural crack damage, and make case-based decisions on the repair of civil aircraft bulkhead structural cracks.

2. The civil aircraft frame structure crack repair decision-making method integrating physical information and case data as described in claim 1, characterized in that, Step 1 involves integrating and learning from current mainstream aircraft frame structure crack repair case data from manufacturers, operators, and maintenance units, and using SVM to perform independence and correlation analysis between indicators, retaining representative indicators, eliminating redundant data, and constructing a civil aircraft frame structure crack repair case library containing structured data, text data, and image data. The data learned from manufacturers is mainly structured data, which consists of various initial data of civil aircraft determined by design and manufacturing, including: design and manufacturing data, geometric and dimensional data, flight performance data, structural health monitoring data, material property data, damage tolerance and fatigue analysis data, and technical publications. The data learned from the operator includes structured data, image data, and text data. It is data generated by the operator during the operation of the aircraft, including: aircraft operation data, aircraft repair records, specific damage information, damage location information, damage discovery records, and damage location images. The data learned from the maintenance unit is mainly text data, which involves specific repair information on cracks and damage to the frame structure, including: description of repair methods, specific repair plans, repair materials, adhesive bonding process, surface treatment and protective measures, and inspection requirements after repair.

3. The civil aircraft frame structure crack repair decision-making method integrating physical information and case data according to claim 1, characterized in that, In step 2, based on the civil aircraft frame structure, select crack tensile test specimens of the same material, simulate the fatigue load spectrum in actual operation, conduct constant amplitude fatigue crack propagation experiments, record the crack length and the corresponding load cycle number, and record 200 data points for each experiment.

4. The civil aircraft bulkhead structure crack repair decision-making method according to claim 1, characterized in that, In step 4, a BERT neural network model is established. The model is initialized using the bert-base-chinese pre-trained weights provided by Google. During the training phase, an MLM training task is constructed using text data from civil aircraft frame structure crack repair cases. The model is then fine-tuned to obtain a language model with characteristics of the civil aircraft frame structure crack repair domain. During the inference phase, the BERT model is used to vectorize the input text and extract features from the case text data. The MLM training task is an unsupervised learning task. The model replaces the segmentation units in the text with a probability of 15%. 80% of the units are replaced with masks, 10% of the units are replaced with any other unit, and the remaining 10% remain unchanged.

5. The civil aircraft frame structure crack repair decision-making method integrating physical information and case data according to claim 1, characterized in that, In step 5, a ViT neural network model is established. The model is initialized using training weights obtained from training on the image dataset of crack detection in the civil aircraft frame structure. During the inference phase, the ViT model is used to extract multi-level features from the input image and extract features from the case image data.

6. The civil aircraft frame structure crack repair decision-making method integrating physical information and case data according to claim 1, characterized in that, In step 6, a twin architecture model with two branch networks is established. Each branch network consists of three neural network models: PINN, BERT, and ViT. The two branches are used to perform multimodal and multidimensional processing and analysis on new cases of crack damage in the bulkhead structure and cases in the historical case library, respectively. The outputs of the two branches are then subjected to similarity analysis. Each new case of crack damage in the bulkhead structure is compared with cases in the historical case library to obtain the repair plan for the most similar case, and case-based decision-making for the repair of cracks in the bulkhead structure of civil aircraft is carried out.

Images