Aircraft assembly quality defect repair scheme generation method based on deep learning

Abstract

An aircraft assembly quality defect repair scheme generation method based on deep learning comprises the following steps: constructing a text classification network model in an offline stage, training by using historical fault feature vectors, extracting keyword features in modular disposal operations, and generating a repair scheme for aircraft assembly quality defects according to the correlation between a text and semantic coordinates of each disposal operation; according to the method, candidate processing operations with relatively high relevancy are screened out to form a modular repair scheme manual, after preprocessing and word segmentation are carried out according to airplane quality problem form data in an online stage, a fault category is predicted through a trained text classification network model, and keyword features with relatively high recommendation relevancy are retrieved and obtained; and outputting a corresponding recommended disposal operation. According to the invention, quality fault repair is carried out through a quality defect diagnosis model and a knowledge mining and disposal scheme generation algorithm.

Description

Technical Field

[0001] The present invention relates to a technology in the field of aircraft manufacturing, and specifically to a method for generating a repair plan for aircraft assembly quality defects based on deep learning. Background Art

[0002] As aircraft assembly quality defects become more complex and diverse, the efficiency and accuracy of aircraft assembly quality defect diagnosis are also required to be higher. Traditional analysis methods mostly focus on a single device or process, rely on the type and reliability of measurement data, and have a small scope of application. Traditional diagnostic processes rely on experience, and there is a large amount of aircraft assembly text data carrying professional knowledge in the civil aircraft manufacturing field to be mined. When the types and scale of quality data increase rapidly, the application of traditional methods is limited. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a method for generating a repair plan for aircraft assembly quality defects based on deep learning. The method structures the text data according to the actual quality defect form data in the aircraft assembly process by combining natural language processing and deep learning techniques, and repairs quality faults through a quality defect diagnosis model and a knowledge mining and disposal plan generation algorithm.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a method for generating a repair plan for aircraft assembly quality defects based on deep learning. In an offline stage, a text classification network model is constructed, and training is performed with historical fault feature vectors. At the same time, keyword features in modular disposal operations are extracted. According to the correlation between the text and the semantic coordinates of each disposal operation, candidate disposal operations with a large correlation are screened out to form a modular repair plan manual. In an online stage, after preprocessing and word segmentation according to aircraft quality problem form data, the fault category is predicted by the trained text classification network model, and then keyword features with a high recommendation correlation are retrieved, and the corresponding recommended disposal operations are output.

[0006] The invention relates to an aircraft assembly quality defect repair system based on deep learning, comprising: a data processing and quality defect classification rule establishment unit, a text classification and model training and prediction unit, a modular disposal operation manual construction and knowledge mining unit, and a complete disposal plan generation unit, wherein: the data processing and quality defect classification rule establishment unit performs segmentation and pruning processing on a long Chinese text according to the Jieba precise pattern word segmentation and word2vec neural network algorithm to obtain a Chinese vocabulary list in vector form, and refers to historical case data, considers the characteristics of various quality defects, classifies all quality defects according to different dimensions, and establishes quality defect classification rules, and the text classification and model training and prediction unit selects a convolutional neural network (CNN) model to extract local semantic features, and selects a long short-term memory (LSTM) The neural network model extracts text context information, constructs a CNN-LSTM neural network model and sets relevant hyperparameters, trains to obtain parameters that achieve the best classification effect, effectively extracts the semantic features and distance features of the text, and thus realizes the diagnosis of new quality problems. The modular disposal operation manual construction and knowledge mining unit summarizes and processes all common quality defect disposal operation methods, and uses the LSA method to extract keyword features of deviation descriptions, and then mines the potential relationship between deviation description feature vocabulary and different disposal solutions, constructs a deviation-disposal knowledge base, and thus obtains disposal solution recommendations with higher correlation, and generates a complete disposal solution unit. According to the keywords of the disposal operation, the modular repair solution manual and the corresponding processing, the disposal operation keywords with higher correlation and their corresponding complete disposal operations are retrieved. Technical Effects

[0007] Compared with the prior art, the present invention uses a quality defect diagnosis model to clean and preprocess mixed data mainly consisting of long Chinese texts, updates a proprietary dictionary in the aircraft manufacturing field, and converts raw text data into a vocabulary list with a high accuracy rate; establishes a deep text classification model to fully extract the semantic features and distance features of the text, and diagnoses aircraft assembly quality problems with a high accuracy rate; the database is iteratively updated with the increase of forms and the feedback from operators, which can assist front-line workers in diagnosing quality problems and improve the accuracy of filling out forms on the one hand, and also lay the foundation for subsequent repair plan decisions on the other hand. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a flow chart of quality defect diagnosis and repair in the present invention;

[0009] Figure 2 It is a schematic diagram of the system of the present invention;

[0010] Figure 3 Generate a flow chart for the embodiment handling opinion. DETAILED DESCRIPTION

[0011] like Figure 1 As shown, a method for generating an aircraft assembly quality defect repair plan based on deep learning based on the above system includes the following steps:

[0012] Step 1: Collect and pre-process the aircraft quality problem form data to obtain a list of Chinese words in vector form, including:

[0013] 1.1) Perform feature pruning on the aircraft quality problem form data from an aircraft manufacturing company, and only retain the attributes related to the deviation and disposal information of quality defects. Specifically, first delete the incomplete or repeated forms, and further compress the data dimensions based on the idea of ​​attribute simplification, and only retain the attributes related to the deviation and disposal information of quality problems, and perform feature pruning on irrelevant attribute information such as flight number, drawing number, and submission time.

[0014] 1.2) Jieba precise mode word segmentation is used to segment a long Chinese text into a vocabulary list and update the aircraft quality defect word segmentation vocabulary. Specifically, the cleaned text to be processed is input, and the general preset dictionary and the proprietary dictionary for the specific situation of the aircraft assembly process are loaded to perform word segmentation in precise mode. For words not in the dictionary, the new word discovery function is enabled. Finally, the word segmentation list of the entire text is obtained.

[0015] 1.3) Use the word2vec neural network algorithm to train the text and convert the Chinese vocabulary list into a vector form. The value on each dimension of the vector represents the correlation level of each word on this dimension. Specifically, the continuous bag of words (CBOW) model and hierarchical softmax training method are used, with the cross entropy loss function as the objective function, and the dimension of the word vector is set to 100 dimensions. Set the training window parameter to 5, that is, when a certain word is used as the output, the 5 words before and after it are used as input. Other parameters include the minimum participating word frequency parameter of 2, the sampling threshold parameter of 0.001, etc. After the model is trained to convergence, the parameter matrix W from the input layer to the hidden layer is stored as a weight matrix, and multiplied with the input vocabulary to realize the vectorized representation of all Chinese words.

[0016] Step 2: Establish quality defect classification rules: Perform manual statistical analysis based on the FRR form of historical case data to identify high-incidence points of quality defects and propose targeted improvement measures; Based on the existing form with accurately marked non-conforming codes, manually consider the characteristics of various quality defects and classify all quality defects into 13 categories according to different dimensions. After the formulated classification rules are evaluated and confirmed by experts, all sample data can be classified;

[0017] The historical case data mentioned refers to: a total of 17,971 FRR data of a certain aircraft manufacturing company from July 2020 to June 2022.

[0018] The 13 categories specifically include: external hook damage, dents, abrasions, scratches, corrosion on the appearance, pits on the appearance, missing corners and defects on the body, other surface problems, hole problems, problems with the fit of two or more parts, fastener installation, functional problems, configuration, deviation of a single part, sealing, and others.

[0019] Step 3: Build a text classification network model and set hyperparameters: Build a convolutional neural network (CNN) model to extract local semantic features and a LSTM model to extract contextual information of the text. Then, connect these two deep learning neural networks and add input and output layers to simultaneously extract the distance features and semantic features of the text to improve the accuracy of text classification. Specifically, the number of hidden layer units H is set to 256, the word vector dimension d is set to 100, the filter window size m is set to 3, 4, 5, the number of convolution kernels T is set to 200, the Dropout ratio p is set to 0.5, the iteration method batch_size is set to 64, the number of iterations epoch is set to 800, and the sentence length seq_length is set to 300.

[0020] The loss function of the text classification network model is L = -∑ k t k log y k , where: the vector output of each data is converted into a set of decimals between 0 and 1, representing the probability that the data corresponds to each category y = softmax(z) = softmax(W T x+b), z j is the output of the fully connected layer in the previous stage, and K is the number of categories. Finally, the one with the greatest possibility is selected as the classification result of the data.

[0021] Since there are many professional long terms in the aircraft FRR form, it is necessary to construct n-gram grammatical features. For example, it is necessary to extract the local feature keyword "left wing upper wall panel" from "left wing", "upper" and "wall panel". Therefore, the CNN model can better extract local semantic features. On the other hand, the context of a deviation description has a strong correlation. For example, the first half of the sentence describes the fault location, and the second half describes the deviation phenomenon of the location. Therefore, the LSTM model can better extract the context information of the text and avoid the dimension disappearance or explosion problem caused by the traditional RNN model. Since convolution and pooling operations will lose the order and position information of the words in the text sequence, it is difficult to capture the context information in the text sequence. Therefore, CNN is suitable for processing spatial data and cannot be directly applied to text data. Therefore, it is spliced ​​after the LSTM unit to retain the distance feature of the text. By splicing these two deep learning neural networks together, and adding the input and output layers, the distance features and semantic features of the text can be extracted at the same time, improving the accuracy of text classification, which will help to improve the efficiency and accuracy of defect repair plan generation in the future.

[0022] Step 4: Model training and prediction: After deleting incomplete or repeated forms, the remaining 17,965 valid data are randomly divided into training set, test set and validation set in a ratio of 10:3:3, and stratified random sampling is performed according to the proportion of label categories obtained in the second step. The training set is used to train the text classification network model, and the validation set is used to verify the classification effect of the text classification network model, and the model parameters with the best prediction effect are saved as the final fault classification diagnosis model.

[0023] In this embodiment, the part of the model with the highest prediction accuracy during the 800 iterations is preferably used.

[0024] According to the actual quality problem form data of the aircraft manufacturing enterprise, the training set, test set and validation set are divided in proportion. According to the principle of the LSTM-CNN deep learning text classification model, the network model is built based on the Tensorflow framework. The hyperparameters and parameters to be trained of the model are initialized, including the number of hidden nodes in each layer, the learning rate, the number of iterations, the weight matrix, and the bias parameter. The network parameters are optimized using the stochastic gradient descent algorithm, and the cross entropy function is used as the loss function for model training. The model is trained using the training set and the validation set is used to evaluate the performance of the network model. If the loss function values ​​of the training set and the validation set converge, the optimal model at this time is saved, and the optimal model is used to predict the test set data. The prediction accuracy is used as the final evaluation indicator of the model effect. The database will also be iteratively updated with the increase of forms and the feedback of operators. On the one hand, it can assist front-line workers in diagnosing quality problems and improve the accuracy of filling in, and on the other hand, it also lays the foundation for subsequent repair plan decisions.

[0025] Step 5: Construct a modular repair solution manual: Relying on various disposal specification systems under the aircraft technical system, manually summarize all common disposal operation methods and compile them into a "Modular Disposal Operation Document", organize them in a form, and extract keyword features to quickly locate the corresponding modular repair solution.

[0026] The keyword features include: grinding, deburring, damage, depth, surface roughness, paint layer, etc. Since the disposal manual is in an unstructured text format, it is organized in a form and keyword features are extracted. When performing subsequent knowledge mining and repair plan recommendations, the corresponding modular disposal operations can be quickly located based on these keywords.

[0027] Step 6: Knowledge mining: Use the LSA method to first extract the keyword features of the deviation description in the FRR form, then mine the potential relationship between the deviation description feature vocabulary and different disposal solutions, and build a deviation-disposal knowledge base. Then calculate the keyword combination with a higher correlation based on the deviation description of the new problem area, and correspond it with the modular repair solution manual formed in the fifth step, retrieve the original text of the repair solution corresponding to the disposal operation keyword and make recommendations, including:

[0028] 6.1) Document feature extraction: Count the frequency of professional vocabulary in the historical FRR form, use semantic structure to represent the relevance of words and texts, use Textrank algorithm to extract keywords from all deviations and disposal words generated after word segmentation, and obtain the weight of keywords Among them: the weight term wji represents the importance of the link between two words. According to the formula, the weight of each node is iteratively calculated until convergence, and the weight can be used to represent the importance of the keyword.

[0029] 6.2) Constructing the latent semantic space: The singular value decomposition (SVD) method is used to construct the latent semantic space of all historical fault rejection reports (FRR). Specifically, according to the deviation feature-treatment plan co-occurrence matrix X, the SVD function in the numpy.lingal module is called to perform singular value decomposition on the matrix X to obtain orthogonal matrices U, V, and singular values ​​S; the first k largest singular values ​​S are taken out. k and the corresponding left and right singular vectors U k , V k ; Then expand the k singular values ​​S into a diagonal matrix S k ; Finally, the coordinate vector US of the deviation description and disposal operation in the latent semantic space is calculated 1 / 2 With VS 1 / 2 .

[0030] 6.3) Perform correlation analysis: Convert the text obtained in step 6.1) into semantic coordinates, and then retrieve the disposal solution terms related to the input deviation feature terms by comparing the correlation between the text and the semantic coordinates of each disposal operation, and filter out several disposal operations with greater correlation as recommendations through a threshold.

[0031] The calculation formula of the correlation of the semantic coordinates is:

[0032] Based on the idea of ​​deep learning, a disposal solution recommendation algorithm is designed based on the latent semantic analysis method to analyze the deviation description text and disposal instruction text of aircraft assembly quality defects. The disposal method in each aircraft assembly fault form is similarly compared to the potential meaning or theme of the deviation description text, that is, it is believed that a certain type of disposal operation corresponds to certain deviation description features. The core is to describe the correlation between text topics and words by extracting the association relationship of text words. After the text data of all FRR forms are segmented by Jieba and custom dictionaries, the importance of each word in the text after segmentation is different. The keywords and weights in the text are extracted by the keyword extraction algorithm, and the text is represented by a set of words and their corresponding weights. For any combination of one or more deviation feature terms, the correlation between the combination and all disposal operations can be calculated, and the operation with the larger correlation can be recommended.

[0033] Step 7: Generate a complete disposal plan in the online stage: Match the disposal operation keywords obtained in the sixth step with the modular repair plan manual formed in the fifth step, retrieve and display the original repair plan corresponding to the disposal operation keywords, specifically: pre-process and segment the new faults input online, then input the data into the diagnosis system, predict the category, and assist in filling in the unqualified code; then retrieve the disposal operation keywords with higher recommended relevance, and display the corresponding complete disposal operations as recommended repair plans.

[0034] According to the assembly quality defect keywords extracted by the automatic identification, the corresponding modular repair solution manual and the original text of the repair solution, a feasible quality defect repair solution recommendation can be made.

[0035] After specific practical experiments, all the words to be processed are first vectorized using the "0-1 method". Assuming that there are V words in total, the word x can be represented as a V-dimensional vector X = (x1, x2, ..., x V ). Each time, the C words before and after word x are taken as input, and the word itself is taken as output. N is the dimension of the dense word vector to be obtained, which is generally 50-500 dimensions. The weight matrix from the input layer to the hidden layer is denoted as W, with a dimension of V×N, and the weight matrix from the input layer to the hidden layer is denoted as W′, with a dimension of N×V.

[0036] The output vocabulary Y=(y1,y2,…,y V ), where the value of each dimension of Y represents the probability that the output word is the corresponding word. The gradient descent method is used to update W and W′. After training, the vector obtained by multiplying each word in the input layer by the matrix W is the required N-dimensional word vector. The objective function of the model is generally the cross entropy loss function Where: w t Indicates the current word, w t-k ,…,w t-1 represents the previous words of word t, w t+1 ,…,w t+k Indicates the following words of word t.

[0037] The model obtains the vocabulary vector Y by calculation. The value of each dimension of the vector is the probability that it is equal to the vocabulary. Therefore, the softmax multi-classifier is usually used to normalize the prediction results and convert the probability value into a specific prediction vocabulary, specifically: Where: y i is the value of each dimension of vector Y, that is, the probability of predicting each word i. Vector Y = (y1, y2, ..., y V ), where: y = b + Uh (w t-k ,w t-k+1 ,…,w t+k ; W), h function is w t-k ,…,w t+k Find the average, U and b are the parameters of the softmax function.

[0038] As shown in Table 1, the treatment plan obtained by the fault treatment opinion generation algorithm for the aircraft assembly process is compared with the actual treatment plan.

[0039] Table 1 Verification of disposal opinion generation results

[0058] Compared with the existing technology, the matching accuracy of this method reached 68.3%, which greatly improved the efficiency of quality problem handling at the production site and effectively used historical data to accumulate and solve more than 100 key repeated problems.

[0059] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A method for generating a repair plan for aircraft assembly quality defects based on deep learning, characterized in that: By constructing a text classification network model in the offline stage and training it with historical fault feature vectors, the keyword features in the modular handling operations are extracted. According to the correlation between the text and the semantic coordinates of each handling operation, the candidate handling operations with higher relevance are screened out to form a modular repair plan manual. In the online stage, after preprocessing and word segmentation according to the aircraft quality problem form data, the fault category is predicted by the trained text classification network model, and then the keyword features with higher recommendation relevance are retrieved, and the corresponding recommended handling operations are output.

2. The method for generating a repair plan for aircraft assembly quality defects based on deep learning according to claim 1 is characterized in that: include: Step 1: Collect and pre-process the aircraft quality problem form data to obtain a Chinese vocabulary list in vector form; Step 2: Establish quality defect classification rules: Perform manual statistical analysis based on the failure rejection report form of historical case data, identify high-incidence points of quality defects, and propose targeted improvement measures; Based on the existing form with accurately marked non-conforming codes, manually consider the characteristics of various quality defects and classify all quality defects into 13 categories according to different dimensions. After the formulated classification rules are evaluated and confirmed by experts, all sample data can be classified; Step 3: Build a text classification network model and set hyperparameters: Build a convolutional neural network model to extract local semantic features and an LSTM model to extract contextual information of the text. Then, connect these two deep learning neural networks and add input and output layers to extract distance features and semantic features of the text at the same time, so as to improve the accuracy of text classification. Step 4: Model training and prediction; Step 5: Build a modular repair solution manual; Step 6: Knowledge mining: Use the LSA method to first extract the keyword features of the deviation description in the FRR form, then mine the potential relationship between the deviation description feature vocabulary and different treatment plans, build a deviation-treatment knowledge base, and then calculate the keyword combination with a higher correlation based on the deviation description of the new problem area, and correspond it with the modular repair plan manual formed in the fifth step, retrieve the original text of the repair plan corresponding to the treatment operation keyword and make a recommendation; Step 7: Generate a complete disposal plan in the online stage.

3. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2 is characterized in that: The first step specifically includes: 1.1) Perform feature pruning on the aircraft quality problem form data from an aircraft manufacturing company, and only retain the attributes related to the deviation and disposal information of quality defects. Specifically, first delete the incomplete or repeated forms, and further compress the data dimension based on the idea of ​​attribute simplification, and only retain the attributes related to the deviation and disposal information of quality problems, and perform feature pruning on irrelevant attribute information such as flight number, drawing number, and submission time; 1.2) Jieba precise mode word segmentation is used to segment a long Chinese text into a vocabulary list and update the aircraft quality defect word segmentation vocabulary. Specifically, the cleaned text to be processed is input, the general preset dictionary and the proprietary dictionary for the specific situation of the aircraft assembly process are loaded, and the word segmentation in precise mode is performed. For words not in the dictionary, the new word discovery function is enabled, and finally the word segmentation list of the entire text is obtained; 1.3) Use the word2vec neural network algorithm to train the text and convert the Chinese vocabulary list into a vector form. The value on each dimension of the vector represents the correlation level of each word on this dimension. Specifically, the continuous bag of words model and hierarchical softmax training method are used, and the cross entropy loss function is used as the objective function. The dimension of the word vector is set to 100 dimensions, and the training window parameter is set to 5. That is, when a certain word is used as the output, the 5 words before and after it are used as input. Other parameters include the minimum participating word frequency parameter of 2 and the sampling threshold parameter of 0.

001. After the model is trained to convergence, the parameter matrix W from the input layer to the hidden layer is stored as a weight matrix. Multiplying it with the input vocabulary realizes the vectorized representation of all Chinese words.

4. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2 is characterized in that: The 13 categories specifically include: external hook damage, dents, abrasions, scratches, corrosion on the appearance, pits on the appearance, missing corners and defects on the body, other surface problems, hole problems, problems with the fit of two or more parts, fastener installation, functional problems, configuration, deviation of a single part, sealing, and others.

5. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2 is characterized in that: The setting of hyper parameters is as follows: the number of hidden layer units H is set to 256, the word vector dimension d is set to 100, the filter window size m is set to 3, 4, 5, the number of convolution kernels T is set to 200, the Dropout ratio p is set to 0.5, the iteration method batch_size is set to 64, the number of iterations epoch is set to 800, and the sentence length seq_length is set to 300; The loss function of the text classification network model is L = -∑ k t k log y k , where: the vector output of each data is converted into a set of decimals between 0 and 1, representing the probability that the data corresponds to each category y = softmax(z) = softmax(W T x+b), z j is the output of the fully connected layer in the previous stage, K is the number of categories, and the one with the greatest possibility is finally selected as the classification result of the data.

6. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2 is characterized in that: The fourth step is specifically as follows: after deleting incomplete or repeated forms, the remaining 17,965 valid data are randomly divided into a training set, a test set, and a validation set in a ratio of 10:3:3, and stratified random sampling is performed according to the proportion of the label categories obtained in the second step. The training set is used to train the text classification network model, and the validation set is used to verify the classification effect of the text classification network model, and the model parameters with the best prediction effect are saved as the final fault classification diagnosis model.

7. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2 is characterized in that: The fifth step is to manually summarize all common disposal operation methods based on various disposal specification systems under the aircraft technical system, organize them into modular disposal operation documents, organize them in a form, and extract keyword features to quickly locate the corresponding modular repair plan; The keyword features include: polishing, deburring, damage, depth, surface roughness, paint layer, etc. Since the disposal manual is in an unstructured text format, it is organized in a form and keyword features are extracted. When performing subsequent knowledge mining and repair plan recommendations, the corresponding modular disposal operations can be quickly located based on these keywords.

8. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2 is characterized in that: The sixth step specifically includes: 6.1) Document feature extraction: Count the frequency of professional vocabulary in the historical FRR form, use semantic structure to represent the relevance of words and texts, use Textrank algorithm to extract keywords from all deviations and disposal words generated after word segmentation, and obtain the weight of keywords Among them: the weight term wji represents the importance of the link between two words. The weight of each node is iteratively calculated according to the formula until convergence, and the weight can be used to represent the importance of the keyword; 6.2) Constructing the latent semantic space: The singular value decomposition method is used to construct the latent semantic space of all historical fault rejection reports. Specifically, according to the deviation feature-treatment plan co-occurrence matrix X, the SVD function in the numpy.lingal module is called to perform singular value decomposition on the matrix X to obtain orthogonal matrices U, V, and singular values ​​S; the first k largest singular values ​​S are taken out. k and the corresponding left and right singular vectors U k , V k ; Then expand the k singular values ​​S into a diagonal matrix S k ; Finally, the coordinate vector US of the deviation description and disposal operation in the latent semantic space is calculated 1 / 2 With VS 1 / 2 ; 6.3) Perform correlation analysis: Convert the text obtained in step 6.1) into semantic coordinates, and then retrieve the disposal solution terms related to the input deviation feature terms by comparing the correlation between the text and the semantic coordinates of each disposal operation, and filter out several disposal operations with greater correlation as recommendations through a threshold.

9. The method for generating an aircraft assembly quality defect repair plan based on deep learning according to claim 2, characterized in that: The seventh step is specifically: matching the treatment operation keywords obtained in the sixth step with the modular repair solution manual formed in the fifth step, retrieving and displaying the original text of the repair solution corresponding to the treatment operation keywords, specifically: pre-processing and segmenting the newly inputted faults online, inputting the data into the diagnosis system, predicting the category to which it belongs, and assisting in filling in the unqualified code; then retrieving the treatment operation keywords with a high recommended relevance, and displaying the corresponding complete treatment operations as recommended repair solutions; According to the assembly quality defect keywords extracted by the automatic identification, the corresponding modular repair solution manual and the original text of the repair solution, a feasible quality defect repair solution recommendation can be made.

10. A system for implementing the method for generating an aircraft assembly quality defect repair plan based on deep learning as described in any one of claims 1 to 9, characterized in that: include: The data processing and quality defect classification rule establishment unit, the text classification and model training and prediction unit, the modular disposal operation manual construction and knowledge mining unit, and the complete disposal plan generation unit, among which: the data processing and quality defect classification rule establishment unit segments and trims the long Chinese text according to the Jieba precise pattern word segmentation and word2vec neural network algorithm to obtain a Chinese vocabulary list in vector form, and refers to historical case data, considers the characteristics of various quality defects, classifies all quality defects according to different dimensions, and establishes quality defect classification rules. The text classification and model training and prediction unit selects the convolutional neural network model to extract local semantic features, selects the long short-term memory neural network model to extract text context information, and splices and constructs C The NN-LSTM neural network model is built and the relevant hyperparameters are set. The parameters that can achieve the best classification effect are trained to effectively extract the semantic features and distance features of the text, thereby realizing the diagnosis of new quality problems. The modular disposal operation manual construction and knowledge mining unit summarizes and processes all common quality defect disposal operation methods, and uses the LSA method to extract the keyword features of the deviation description, and then mines the potential relationship between the deviation description feature vocabulary and different disposal solutions, constructs a deviation-disposal knowledge base, and thus obtains disposal solution recommendations with higher correlation. The complete disposal solution unit generates the disposal operation keywords and their corresponding complete disposal operations based on the disposal operation keywords, the modular repair solution manual and the corresponding processing.

Images