A method, system, device and medium for reasoning process specifications of aviation parts
Through the aeronautical parts process procedures inference method, using feature similarity and multi-dimensional fusion technology, neural networks and weight gain loss functions are built, which solves the problem of relying on experience in traditional process design, and realizes the rapid and intelligent compilation of part process procedures, which improves the Design consistency and accuracy.
Patent Information
- Application Number
- CN202410978861.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-22
AI Technical Summary
The traditional aviation parts process design process relies on personal experience, resulting in poor consistency, difficulty in optimization, weak standardization, long design cycle, difficult to adapt to changes in the product and manufacturing environment, and difficult to control data accuracy.
By using the aeronautical parts process regulations inference method, the features of parts to be processed are extracted, the characteristics similarity are calculated, the multi-dimensional feature vector fusion is carried out, and the neural network and weight gain triple loss function are constructed to obtain the optimal matching process regulations to realize the parts Rapid preparation and intelligent preparation of process regulations.
The design cycle is shortened, matching accuracy and consistency is improved, important feature information is retained, and the model's distinction ability and overall performance are enhanced.
Smart Images

Figure CN118504694B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent process technology, and in particular to an aviation parts process procedure reasoning method, system, equipment and medium. Background Art
[0002] In the traditional process design process, the preparation of process specifications is too dependent on the personal experience and level of the process designers. The rationality, operability and preparation time of process documents mainly depend on the experience and proficiency of the process personnel; at the same time, the preparation cycle is long, there is a lot of repetitive work, and it is difficult to adapt to changes in products and manufacturing environments.
[0003] However, the internal structure of the aircraft is complex, the parts are of different shapes and in large quantities, and its development process is very complicated. Most of its products are accompanied by the characteristics of multiple varieties and small batches. Due to limited personal experience and different design levels, the designed process documents vary from person to person, with poor consistency, difficult optimization, weak standardization, and difficult inheritance of process design experience. With the transformation and upgrading of the aerospace manufacturing industry, new requirements have been put forward for the preparation of process procedures for complex products such as modern aircraft products.
[0004] It is of great significance to develop process specifications and push relevant knowledge to achieve rapid and intelligent development of part process specifications.
[0005] In the production process of discrete manufacturing workshops, most of them adopt the "multi-variety, small batch" production model, so the workload of compiling part process regulations is large and repetitive.
[0006] For the design and manufacturing of a part, this link can be divided into three major processes: product design-process design-product manufacturing. The process document generated by the process design is an important document to guide the production process and the basis for formulating production plans and heights. Its quality and efficiency directly affect the configuration and optimization of the company's manufacturing resources, product quality, production organization efficiency, product cost and production cycle, etc.
[0007] Due to the continuous development of the current aviation aircraft manufacturing industry, a high degree of informatization and automation has been achieved for the product design and product manufacturing links in the entire process. As for traditional process design, it varies greatly depending on the resources and process habits of the enterprise. Under the same resources and constraints, different process designers may formulate different process regulations. This is a decision-making process with strong empiricism and many influencing factors. Therefore, traditional industrial design has the following defects:
[0008] (1) It is not easy to standardize, has poor consistency, and is overly dependent on the personal experience and level of process designers. Since the traditional process design process is greatly affected by the design ability and design level of process designers, the actual process design process has poor consistency, weak standardization, and difficulty in optimization, which also affects the inheritance and integrated utilization of process design experience. The rationality, operability, and preparation time of process documents mainly depend on the experience and proficiency of process personnel. At the same time, since it is impossible to standardize process design experience and knowledge, it is not convenient to concentrate process experts and process design experience and knowledge for full utilization.
[0009] (2) The design cycle is long and the sensitivity to product changes is weak. For the process design of new parts, even if the parts are similar to the original parts, it is impossible to fully utilize the reusable process specifications between similar parts. Redesign is necessary, which results in a considerable amount of unnecessary design work and makes it difficult to adapt to changes in products and manufacturing environments. Process designers are forced to spend a lot of energy on repetitive and tedious work and lack research on innovative process work.
[0010] (3) Data accuracy cannot be fully controlled. As the demand for component design increases, the process design process involves the processing of large amounts of heterogeneous data, such as graphic information, data information, and text information. At the same time, process design will generate a large amount of process files and process data. The traditional method of processing data in process design lacks standardization and informatization, and is not fully optimized, which is very prone to errors.
[0011] The internal structure of an aircraft is complex, the number of parts is large and diverse, and different overall aircraft designs are accompanied by different parts designs, making the aircraft parts development process very complex and accompanied by the characteristics of multiple varieties and small batches. Nowadays, with the transformation and upgrading of the aviation manufacturing industry, the demand for aircraft parts process design is increasing, and it is more important to realize the rapid and intelligent compilation of parts process specifications for aircraft parts products. Summary of the invention
[0012] In view of the problems of poor inconsistency of part process specifications, long design cycle and inability to fully control data accuracy, the present invention proposes an aviation part process specification reasoning method, system, equipment and medium; firstly, the features of the part to be processed are extracted and data is embedded, and the feature similarity is calculated; then, the part instance obtained by preliminary matching is fused with the part instance feature vector obtained from the part instance library through multi-dimensional feature vector fusion to obtain a fusion vector; finally, a neural network and a weight gain triplet loss function are constructed to obtain the best matching part instance, and the process specification of the best matching instance part is reused; this ensures that important feature information is retained during vector fusion, and improves matching accuracy while shortening the design cycle.
[0013] The specific implementation contents of the present invention are as follows:
[0014] A method for reasoning about process specifications for aviation parts is proposed. First, the features of the parts to be processed are extracted, and the features of the parts to be processed are embedded in data to calculate feature similarity. Then, a part instance is obtained by preliminary matching based on feature similarity, and a multi-dimensional feature vector fusion is performed with the part instance feature vector obtained from the part instance library to obtain a fusion vector. Finally, a neural network and a weight gain triplet loss function are constructed to obtain the optimal matching part instance, and the process specification of the optimal matching instance part is reused.
[0015] In order to better implement the present invention, further, the aviation parts process specification reasoning method specifically includes the following steps:
[0016] Step S1: extract the features of the part to be processed, embed the features of the part to be processed into data, and calculate the similarity of geometric structure features and the similarity of part process features;
[0017] Step S2: Preliminary matching is performed to obtain the part instance based on the similarity of the geometric structure feature and the similarity of the part process feature, and a multi-dimensional feature vector fusion is performed with the part instance feature vector obtained from the part instance library to obtain a fusion vector;
[0018] Step S3: Based on the constructed neural network and weight gain triplet loss function, the similarity score is calculated, the instance parts to be matched first are determined, and the process specifications of the instance parts with the best matching are reused.
[0019] In order to better implement the present invention, further, the step S1 specifically includes the following steps:
[0020] Step S11: collecting data of the part to be processed and extracting features of the part to be processed;
[0021] Step S12: performing one-hot encoding on the extracted features of the parts to be processed to obtain a geometric structure feature data set and a part process feature data set;
[0022] Step S13: constructing a geometric structure feature similarity calculation model according to the geometric structure feature data set, and constructing a part process feature similarity calculation model according to the part process feature data set;
[0023] Step S14: Calculate the geometric structure similarity vector according to the geometric structure feature similarity calculation model, and construct the part process feature similarity calculation model according to the part process feature data set to calculate the part process feature similarity vector.
[0024] In order to better implement the present invention, further, the step S2 specifically includes the following steps:
[0025] Step S21: Preliminary matching is performed to obtain a part instance based on the geometric structure feature similarity and the part process feature similarity, in combination with a set geometric structure similarity threshold and a set process similarity threshold;
[0026] Step S22: performing multi-dimensional feature vector fusion on the part instance obtained by preliminary matching and the part instance feature vector obtained from the part instance library to obtain a fusion vector.
[0027] In order to better implement the present invention, further, the step S22 specifically includes the following steps:
[0028] Step S221: Acquire geometric feature data h from the geometric structure feature data set i , obtain the process feature data h from the part process feature data set j ;
[0029] Step S222: According to the set weight tensor W, multi-dimensional fusion geometric feature data h i With process characteristic data h j , get the multi-dimensional fusion vector x a ;
[0030] Step S223: Call the voting mechanism to fuse the geometric feature data h according to the set weight vector V i With process characteristic data h j , get the voting election fusion vector x b ;
[0031] Step S224: According to the set bias vector, call the activation function to connect the activation fusion vector x a and the fusion vector x b , and obtain the fused feature vector g(x a , x a ).
[0032] In order to better implement the present invention, further, the step S3 specifically includes the following steps:
[0033] Step S31: Construct a neural network model and transform the feature vector g(x a , x a ) is input into the neural network model to obtain the feature vector g(x a , x a ) The data after passing through the neural network f (Anchor i );
[0034] The number of hidden layers of the neural network model is L, the number of nodes in each layer is pl, and the weight between the initial hidden layer and the output layer is W t , the offset vector is b l, the learning rate is γ, the iteration threshold is ε, and the activation function is f(·);
[0035] Step S32: According to the data f (Anchor i ), the obtained positive sample fusion vector, the obtained negative sample fusion vector, the set boundary value α, the set weight gain w1, the set weight gain w2, construct a triplet loss function based on weight gain, and calculate the loss value L triplet ;
[0036] Step S33: According to the loss value L triplet , calculate the gradient of the neural network layer, and modify the parameters of the corresponding neural network layer according to the gradient of the neural network layer to obtain the optimal matching model;
[0037] Step S34: Calculate the similarity score according to the optimal matching model, determine the instance part that is matched first, and reuse the process specification of the instance part that is matched best.
[0038] Based on the above-mentioned aviation parts process specification reasoning method, in order to better realize the present invention, further, an aviation parts process specification reasoning system is proposed, which is used to execute the above-mentioned aviation parts process specification reasoning method; it includes a similarity calculation module, a fusion module, and an optimal matching module;
[0039] The similarity calculation module is used to extract the features of the parts to be processed, embed the features of the parts to be processed into data, and calculate the feature similarity;
[0040] The fusion module is used to obtain a part instance by preliminary matching based on feature similarity, and to perform multi-dimensional feature vector fusion with the part instance feature vector obtained from the part instance library to obtain a fusion vector;
[0041] The optimal matching module is used to construct a neural network and a weight gain triplet loss function to obtain an optimal matching part instance and reuse the process specification of the optimal matching instance part.
[0042] Based on the above-mentioned aviation part process procedure reasoning method, in order to better realize the present invention, further, an electronic device is proposed, including a memory and a processor; the memory is used to store computer programs; when the computer program is executed on the processor, the above-mentioned aviation part process procedure reasoning method is implemented.
[0043] Based on the above-mentioned aviation part process procedure reasoning method, in order to better realize the present invention, further, a computer-readable storage medium is proposed, on which computer instructions are stored; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned aviation part process procedure reasoning method is implemented.
[0044] The present invention has the following beneficial effects:
[0045] (1) The present invention uses a fusion method after a voting mechanism and a triplet loss function based on weight gain to ensure that the algorithm retains important feature information when performing vector fusion, and accelerates the training process through the triplet loss function of weight gain, thereby improving the matching accuracy.
[0046] (2) The present invention adopts a feature fusion algorithm based on multi-dimensional fusion features and voting mechanism to fuse part process features and geometric structure features in a high-dimensional vector space by adopting a multi-dimensional and voting method. While retaining the important feature information of the part instance, each part instance is projected into the same semantic space, thereby improving the matching accuracy.
[0047] (3) The present invention uses a triplet loss function with weight gain to minimize the distance between similar samples and maximize the distance between different samples, so that the model can better distinguish samples of different categories. At the same time, the weight gain is adopted to encourage the model to pay more attention to the distinction of samples, effectively improving the overall performance of the model and the computational training process. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 A schematic flowchart of the aviation parts process specification reasoning method provided by the present invention.
[0049] Figure 2 This is a schematic diagram of the optimal matching process provided by the present invention.
[0050] Figure 3 A schematic flow chart of the aviation parts process specification reasoning method provided by the present invention. DETAILED DESCRIPTION
[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be understood that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, and therefore should not be regarded as limiting the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technical personnel in this field without making creative work are within the scope of protection of the present invention.
[0052] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "disposed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0053] Embodiment 1:
[0054] This embodiment proposes a method for reasoning about a process specification of an aviation part. First, the features of the part to be processed are extracted, and the features of the part to be processed are embedded in data to calculate feature similarity; then, a part instance is obtained by preliminary matching based on feature similarity, and a multi-dimensional feature vector fusion is performed with the part instance feature vector obtained from the part instance library to obtain a fusion vector; finally, a neural network and a weight gain triplet loss function are constructed to obtain the best matching part instance, and the process specification of the best matching instance part is reused. The method for reasoning about a process specification of an aviation part specifically includes the following steps:
[0055] Step S1: extract the features of the part to be processed, embed the features of the part to be processed into data, and calculate the similarity of geometric structure features and part process features.
[0056] The step S1 specifically includes the following steps:
[0057] Step S11: collecting data of the part to be processed and extracting features of the part to be processed;
[0058] Step S12: performing one-hot encoding on the extracted features of the parts to be processed to obtain a geometric structure feature data set and a part process feature data set;
[0059] Step S13: constructing a geometric structure feature similarity calculation model according to the geometric structure feature data set, and constructing a part process feature similarity calculation model according to the part process feature data set;
[0060] Step S14: Calculate the geometric structure similarity vector according to the geometric structure feature similarity calculation model, and construct the part process feature similarity calculation model according to the part process feature data set to calculate the part process feature similarity vector.
[0061] Step S2: Preliminary matching is performed to obtain the part instance based on the similarity of the geometric structure features and the similarity of the part process features, and a multi-dimensional feature vector fusion is performed with the part instance feature vector obtained from the part instance library to obtain a fusion vector.
[0062] The step S2 specifically includes the following steps:
[0063] Step S21: Preliminary matching is performed to obtain a part instance based on the geometric structure feature similarity and the part process feature similarity, in combination with a set geometric structure similarity threshold and a set process similarity threshold;
[0064] Step S22: performing multi-dimensional feature vector fusion on the part instance obtained by preliminary matching and the part instance feature vector obtained from the part instance library to obtain a fusion vector.
[0065] The step S22 specifically includes the following steps:
[0066] Step S221: Acquire geometric feature data h from the geometric structure feature data set i , obtain the process feature data h from the part process feature data set j ;
[0067] Step S222: According to the set weight tensor W, multi-dimensional fusion geometric feature data h i With process characteristic data h j , get the multi-dimensional fusion vector x a ;
[0068] Step S223: Call the voting mechanism to fuse the geometric feature data h according to the set weight vector V i With process characteristic data h j , get the voting election fusion vector x b ;
[0069] Step S224: According to the set bias vector, call the activation function to connect the activation fusion vector x a and the fusion vector x b , and obtain the fused feature vector g(x a , x a ).
[0070] Step S3: Based on the constructed neural network and weight gain triplet loss function, the similarity score is calculated, the instance parts to be matched first are determined, and the process specifications of the instance parts with the best matching are reused.
[0071] The step S3 specifically comprises the following steps:
[0072] Step S31: Construct a neural network model and transform the feature vector g(x a , x a ) is input into the neural network model to obtain the feature vector g(x a , x a ) The data after passing through the neural network f (Anchor i );
[0073] The number of hidden layers of the neural network model is L, the number of nodes in each layer is pl, and the weight between the initial hidden layer and the output layer is W t , the offset vector is b l , the learning rate is γ, the iteration threshold is ε, and the activation function is f(·);
[0074] Step S32: According to the data f (Anchor i ), the obtained positive sample fusion vector, the obtained negative sample fusion vector, the set boundary value α, the set weight gain w1, the set weight gain w2, construct a triplet loss function based on weight gain, and calculate the loss value L triplet ;
[0075] Step S33: According to the loss value L triplet , calculate the gradient of the neural network layer, and modify the parameters of the corresponding neural network layer according to the gradient of the neural network layer to obtain the optimal matching model;
[0076] Step S34: Calculate the similarity score according to the optimal matching model, determine the instance part that is matched first, and reuse the process specification of the instance part that is matched best.
[0077] Working principle: This embodiment first extracts the features of the parts to be processed and embeds the data, and calculates the feature similarity; then, the part instance obtained by preliminary matching is fused with the part instance feature vector obtained from the part instance library in a multi-dimensional manner to obtain a fusion vector; finally, a neural network and a weight gain triplet loss function are constructed to obtain the optimal matching part instance, and the process procedures of the optimal matching instance parts are reused; this ensures that important feature information is retained during vector fusion, thereby shortening the design cycle and improving matching accuracy.
[0078] Embodiment 2:
[0079] This embodiment is based on the above embodiment 1. Figure 1 , Figure 2 , Figure 3 As shown, a specific embodiment is described in detail.
[0080] Step S1: collect part data, obtain data on corresponding part process features and geometric structure features from various sources, and construct a similarity calculation model based on part process features and part geometric structure features; calculate the similarity based on part geometric structure features and process features respectively according to part process features and geometric structure features, perform preliminary matching, and screen out similar part instances;
[0081] The process features of the parts in this embodiment include: rolling wave, welding type, surface treatment, inspection type, etc.;
[0082] The geometric structure features of the parts in this embodiment include: material state, rough material size, ovality, inner diameter, outer diameter, bending diameter, etc.
[0083] Step S2: perform part feature vector fusion; after preliminary matching of similarity vectors based on part geometric structure features and part process features in step S1, a tensor fused with two features is obtained by feature vector fusion, and a fused feature vector is extracted by multi-dimensional fusion features and voting mechanism to provide data for subsequent optimal similarity matching;
[0084] Step S3: similarity optimal matching; input the fusion tensor obtained in step S2 into the deep neural network of the optimal matching algorithm based on the weighted gain triple loss function, output a predicted similarity score, and select the best matching instance part according to the final similarity score; calculate a predicted similarity score according to the weighted gain triple loss function, and select the best matching instance part according to the obtained similarity score;
[0085] Determine the matching part instance; based on the calculated final similarity with each part, comprehensively analyze the similarity of the parts corresponding to each instance in the process specification instance library, realize the optimal matching calculation of the similarity between the parts to be processed and the parts corresponding to the process specification instance, and optimally match the process specification instance with the highest similarity for the parts to be processed, thereby completing the case-based aviation part process specification reasoning;
[0086] Reuse instance process specifications; based on the best matching parts obtained, select the process specifications of the instance parts with the highest matching degree for reuse.
[0087] The feature-based similarity calculation model includes the following steps:
[0088] Step 1) Perform one-hot encoding on the geometric structural features of the parts, such as material state, rough material size, ovality, inner diameter, outer diameter, bending diameter, etc., and process features such as rolling wave, welding type, surface treatment, inspection type, etc. Each feature of the part can be uniquely represented by one-hot encoding to obtain the processed data set:
[0089] ; ;
[0090] Among them, S1 is the geometric structure feature data set of the part, S2 is the process structure data set of the part, and X ij is the unique hot vector of the jth part in the ith data set, and m is the maximum number of parts in the data set.
[0091] Step 2) Establish a similarity calculation model for the features of the parts. The similarity used is the Euclidean distance, which is used to calculate the absolute distance between two points in the D-dimensional space. Assume that there are currently two D-dimensional vectors and Among them, X1 and X2 belong to S1 but are two parts. The geometric structure similarity between the parts can be expressed by the following Euclidean formula:
[0092]
[0093] Among them, X1 and X2 are two D-dimensional vectors, X1 and X2 belong to S1 but belong to two parts. x ij is a feature belonging to part i in the vector, is the Euclidean distance between two D-dimensional vectors;
[0094] Among them, K is the calculation parameter, which represents a 1-D positive integer, X1 and X2 are two D-dimensional vectors, x ij is a feature in the vector, is the Euclidean distance between two D-dimensional vectors; the Euclidean distance is the similarity between two vectors.
[0095] according to Formula, the geometric structure similarity can be calculated Similarity to process For two parts, the geometric feature vectors and process feature vectors of the two parts can be obtained respectively. Then the geometric structure similarity between the two parts is the Euclidean distance between the geometric feature vectors of the two parts, and the process similarity is the Euclidean distance between the process feature vectors of the two parts.
[0096] Step 3), based on the geometric structure similarity and process similarity between the OEM parts and the instance parts obtained in step 2), the parts can be preliminarily matched to screen out roughly similar part instances. The screening formula is as follows:
[0097]
[0098] Among them, α1 is the geometric structure similarity distance threshold, α2 is the process similarity distance threshold, && represents the logical relationship "and", and the threshold can be obtained through experiments. Only when the calculated similarity meets the requirement of being less than the similarity distance threshold, the corresponding part instance is roughly similar to the part to be processed.
[0099] The example-based reasoning method for process specification of aviation parts, the feature fusion algorithm of parts based on multi-dimensional fusion features and voting mechanism includes the following steps:
[0100] Step 1) Obtain the geometric structure feature data h of the same part based on the obtained geometric structure feature data set and process feature data set. i and process characteristic data h j , and h i and h j To perform multi-dimensional fusion features, the formula is as follows:
[0101]
[0102] Among them, x a h i and h j The vector after multi-dimensional fusion features, ; is a weight tensor, which represents K D×D matrices, where D is the number of rows and columns of the matrix. h is calculated by the weights in the matrix. i and h j The fusion features are extracted and weights are assigned to each fusion feature; K is a hyperparameter that controls the number of interaction similarity scores generated by the model for each similarity vector pair. Represents a domain;
[0103] Step 2), for the geometric structure feature data h i and process characteristic data h j The feature adopts a voting election mechanism, and the formula is as follows:
[0104]
[0105] Among them, x b is the fusion vector after the voting mechanism. ; [] indicates a cascade operation, indicating a vector h i and h j Merge on dimension 0; is the weight vector;
[0106] Step 3) Perform a connection activation operation based on the two fusion vectors of step 1) and step 2). The specific formula is as follows:
[0107]
[0108] Where b is the bias vector, ; is an activation function; through multi-dimensional fusion features and voting mechanism, a fused feature vector can be finally obtained , .
[0109] The example-based reasoning method for aviation parts process specification, the optimal matching algorithm based on the weighted gain triple loss function comprises the following steps:
[0110] Step 1), according to the obtained fusion vector, input into the number of hidden layers L and the number of nodes in each layer p l (l=1,2,...L), initialize the weights W between the hidden layer and the output layer t , offset vector b l (l=1,2,...L), the learning rate is γ, the iteration threshold is ε, and the activation function of the neural network is f(∙) which is the ReLu function;
[0111] Step 2), since the optimal matching algorithm is to perform the optimal matching of the similarity of the parts to be processed, it is impossible to obtain the similarity between the parts to be processed and the corresponding part instances. Therefore, a triplet loss function based on weight gain is adopted, and the formula is as follows:
[0112]
[0113] in, is the triple loss function; N is the number of samples in a training batch, Represents the data after the fusion vector of the processed parts passes through the neural network; Represents the positive sample Positive fusion vector, which belongs to the same category as the processed parts; Represents the negative sample Negative fusion vector, which belongs to a different category from the processed parts; is the L2 norm, used to calculate the length of the vector; α is the margin, used to process the distance between the fusion vector of the part and the fusion vector of the negative sample, usually a positive value; w1 and w2 are weight gains, used to control The obtained length is guaranteed The importance of both parts increases the convergence speed of the algorithm and improves the accuracy;
[0114] Step 3), according to the loss value obtained in step 2), that is, , calculate the gradient of each layer; modify the parameters of each layer network through the obtained gradient to ensure the accuracy of the algorithm;
[0115] Step 4) If the training process reaches the upper limit of the number of iterations or meets the error constraint, the training process ends. Otherwise, continue training.
[0116] The other parts of this embodiment are the same as those of the above-mentioned embodiment 1, and thus will not be described in detail.
[0117] Embodiment 3:
[0118] Based on any one of the above-mentioned embodiments 1-2, this embodiment proposes an aviation part process specification reasoning system for executing the above-mentioned aviation part process specification reasoning method; including a similarity calculation module, a fusion module, and an optimal matching module;
[0119] The similarity calculation module is used to extract the features of the parts to be processed, embed the features of the parts to be processed into data, and calculate the feature similarity;
[0120] The fusion module is used to obtain a part instance by preliminary matching based on feature similarity, and to perform multi-dimensional feature vector fusion with the part instance feature vector obtained from the part instance library to obtain a fusion vector;
[0121] The optimal matching module is used to construct a neural network and a weight gain triplet loss function to obtain an optimal matching part instance and reuse the process specification of the optimal matching instance part.
[0122] This embodiment also proposes an electronic device, including a memory and a processor; the memory is used to store a computer program; when the computer program is executed on the processor, the above-mentioned aviation part process procedure reasoning method is implemented.
[0123] This embodiment also proposes a computer-readable storage medium, on which computer instructions are stored; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned aviation part process procedure reasoning method is implemented.
[0124] Working principle: Figure 1 As shown, it includes a data module, a similarity calculation module, an optimal matching module and an interface module;
[0125] The data module includes the digital model information of the processed parts and the process specification knowledge base;
[0126] The processing part digital model information is extracted from the part digital model to be processed, and is used to convert the geometric structure characteristics and process characteristic data of the part to be processed that can be accepted by the process specification reasoning algorithm; the part digital model to be processed contains the geometric structure characteristic data and process characteristic data of the part; the part geometric characteristics include the part outer diameter, part wall thickness, part bending radius, and guide part length; the part process characteristic data includes the part welding category, part surface treatment, and part air tightness;
[0127] The process specification knowledge base stores process specification knowledge, which is used to generate process parameters and process flow; the process specification knowledge includes part process feature knowledge, part geometric structure feature knowledge, process specification template knowledge, process flow decision knowledge, tooling tool selection knowledge, and process parameter decision knowledge; the part process feature knowledge and part geometric structure feature knowledge include the process features and geometric structure features of the historical products of the part; the process specification template knowledge includes composite template knowledge and catheter template knowledge, which stores the content and attributes used to describe different types of templates; the process flow decision knowledge includes the record and rule base of decision judgments on the process flow; the tooling tool selection knowledge includes the method rules for determining tooling tools according to part features and process features; the process parameter decision knowledge includes the reasoning decision rules for process parameters;
[0128] The similarity calculation module performs similarity calculation on the processing parts and the part instances, and is used to preliminarily compare the similarity between the processing parts and the instances in the process specification knowledge base;
[0129] The optimal matching module fuses the process feature similarity obtained by the similarity calculation module with the part geometric structure feature similarity to obtain a fusion tensor; the fusion tensor is a vector containing important features of the process feature similarity and the part geometric structure feature similarity, which is used to input into the artificial neural network of the similarity optimal matching algorithm, and performs similarity calculation by matching the part instance features obtained by the initial screening of the continued feature fusion one by one, outputs a predicted similarity score, and selects the best matching instance part according to the optimal similarity score. The artificial neural network is a deep neural network, which is used to calculate and output the final similarity score;
[0130] The interface module can provide the demonstration model to system administrators, process designers, production operators and other operators for use; the system administrator has the highest authority and can manage all the underlying graph database information of the system, including personnel information, permissions, databases, process resources and other information, and perform operations such as adding and deleting to ensure data updates; the process designers can view data such as process knowledge, and use the underlying data to implement related functions for process file generation; the process auditors can use related functions for manufacturing process file management; the production operators receive manufacturing process files to implement paperless production and processing of products.
[0131] The other parts of this embodiment are the same as any one of the above-mentioned embodiments 1-2, so they will not be repeated here.
[0132] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for reasoning about process specifications for aviation parts, characterized in that: First, the features of the parts to be processed are extracted, and the features of the parts to be processed are embedded in data to calculate the feature similarity; then, the part instance is obtained by preliminary matching based on the feature similarity, and the multi-dimensional feature vector is fused with the part instance feature vector obtained from the part instance library to obtain the fusion vector; finally, a neural network and a weight gain triplet loss function are constructed to obtain the optimal matching part instance, and the process specification of the optimal matching instance part is reused; The aviation parts process specification reasoning method specifically comprises the following steps: Step S1: extract the features of the part to be processed, embed the features of the part to be processed into data, and calculate the similarity of geometric structure features and the similarity of part process features; Step S2: Preliminary matching is performed to obtain the part instance based on the similarity of the geometric structure feature and the similarity of the part process feature, and a multi-dimensional feature vector fusion is performed with the part instance feature vector obtained from the part instance library to obtain a fusion vector; Step S3: Calculate the similarity score based on the constructed neural network and weight gain triplet loss function, determine the instance parts to be matched first, and reuse the process specification of the instance parts with the best matching; The step S1 specifically includes the following steps: Step S11: collecting data of the part to be processed and extracting features of the part to be processed; Step S12: performing one-hot encoding on the extracted features of the parts to be processed to obtain a geometric structure feature data set and a part process feature data set; Step S13: constructing a geometric structure feature similarity calculation model according to the geometric structure feature data set, and constructing a part process feature similarity calculation model according to the part process feature data set; Step S14: calculating the geometric structure similarity vector according to the geometric structure feature similarity calculation model, and constructing the part process feature similarity calculation model according to the part process feature data set to calculate the part process feature similarity vector; The step S2 specifically includes the following steps: Step S21: Preliminary matching is performed to obtain a part instance based on the geometric structure feature similarity and the part process feature similarity, in combination with a set geometric structure similarity threshold and a set process similarity threshold; Step S22: performing multi-dimensional feature vector fusion on the part instance obtained by preliminary matching and the part instance feature vector obtained from the part instance library to obtain a fusion vector; The step S22 specifically includes the following steps: Step S221: Acquire geometric feature data h from the geometric structure feature data set i , obtain the process feature data h from the part process feature data set j ; Step S222: According to the set weight tensor W, multi-dimensional fusion geometric feature data h i With process characteristic data h j , get the multi-dimensional fusion vector x a ; Step S223: Call the voting mechanism to fuse the geometric feature data h according to the set weight vector V i With process characteristic data h j , get the voting election fusion vector x b ; Step S224: According to the set bias vector, call the activation function to connect the activation fusion vector x a and the fusion vector x b , and obtain the fused feature vector g(x a , x a ); The step S3 specifically comprises the following steps: Step S31: Construct a neural network model and transform the feature vector g(x a , x a ) is input into the neural network model to obtain the feature vector g(x a , x a ) The data after passing through the neural network f (Anchor i ); The number of hidden layers of the neural network model is L, the number of nodes in each layer is pl, and the weight between the initial hidden layer and the output layer is W t , the offset vector is b l , the learning rate is γ, and the iteration threshold is , the activation function is f(·); Step S32: According to the data f (Anchor i ), the obtained positive sample fusion vector, the obtained negative sample fusion vector, the set boundary value α, the set weight gain w1, the set weight gain w2, construct a triplet loss function based on weight gain, and calculate the loss value L triplet ; Step S33: According to the loss value L triplet , calculate the gradient of the neural network layer, and modify the parameters of the corresponding neural network layer according to the gradient of the neural network layer to obtain the optimal matching model; Step S34: Calculate the similarity score according to the optimal matching model, determine the instance part that is matched first, and reuse the process specification of the instance part that is matched best.
2. An aviation parts process specification reasoning system, used to execute the aviation parts process specification reasoning method according to claim 1; characterized in that: Including similarity calculation module, fusion module and optimal matching module; The similarity calculation module is used to extract the features of the parts to be processed, embed the features of the parts to be processed into data, and calculate the feature similarity; The fusion module is used to obtain a part instance by preliminary matching based on feature similarity, and to perform multi-dimensional feature vector fusion with the part instance feature vector obtained from the part instance library to obtain a fusion vector; The optimal matching module is used to construct a neural network and a weight gain triplet loss function to obtain an optimal matching part instance and reuse the process specification of the optimal matching instance part.
3. An electronic device, characterized in that: It comprises a memory and a processor; the memory is used to store a computer program; when the computer program is executed on the processor, the aviation parts process procedure reasoning method as claimed in claim 1 is implemented.
4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 3, the aviation parts process procedure reasoning method as described in claim 1 is implemented.
Citation Information
Cited By
Process scheme generation method based on image data and knowledge graph
CN120849645A
