A manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning

Abstract

The application discloses a manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning. First, based on the part instance model and the related XML file, the part information is extracted, the features and the manufacturing process sequence are obtained, and the initial data set is formed. Second, the PrefixSpan algorithm is used for rule mining, and the manufacturing process knowledge base is constructed. Then, taking this as the training sample, a deep structured semantic model based on double-channel attention and self-supervised learning is constructed. When the part feature data and the frequent manufacturing process sequence are input into the model, data updating is carried out through the double-channel attention mechanism, and representation learning is carried out based on self-supervised learning. The updated part feature data and the frequent manufacturing process sequence are subjected to similarity calculation, and the frequent sequence is sorted according to the similarity and then recommended. The application proposes the application of knowledge driving in the manufacturing process sequence recommendation, improves the intelligentization and efficiency of process compilation, and has important significance for improving the process design level of the aviation manufacturing industry.

Description

Technical Field

[0001] This invention relates to the field of manufacturing process design, and more specifically to a method for recommending manufacturing process sequences based on knowledge mining and self-supervised learning. Background Technology

[0002] With the continuous development of the aerospace manufacturing industry, both product design and manufacturing processes have achieved a high degree of informatization and automation. Process design is a crucial link in ensuring product quality and production efficiency in modern aerospace manufacturing. Simultaneously, with the continuous advancement of aerospace technology, the variety and complexity of parts are increasing daily, and traditional process design relying on the experience of process engineers cannot solve the "information overload" phenomenon. Knowledge-driven process design, as a new design model within process design, can effectively address the "information overload" phenomenon and improve the intelligence of process design, and has gradually emerged in the aerospace manufacturing industry.

[0003] Current research has focused on organizing and summarizing existing process data to systematize and structure discrete process information, forming a shared and reusable knowledge base. This aims to improve the efficiency and quality of process design and ensure the consistency and stability of process design. However, the construction of this knowledge base involves high computational complexity and low efficiency. Some studies remain at the level of extracting new, shallow features from parts without delving into the acquisition of important information from deep vector spaces. Furthermore, some studies fail to consider the skewed distribution of data when recommending sequences, making it difficult for sequence matching algorithms to learn the distribution of all data effectively, thus hindering effective sequence matching for datasets with small sample sizes. Summary of the Invention

[0004] Purpose of the invention: This invention proposes a manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning, which achieves highly accurate manufacturing process sequence recommendation.

[0005] Technical solution: The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning described in this invention includes the following steps:

[0006] (1) Based on the part instance model information and related XML file information, the instance information is read through CATIA secondary development and XPath, and the part features and manufacturing process sequences of the instance model and related XML files are collected to construct the initial manufacturing dataset;

[0007] (2) Based on the Prefixspan sequence pattern mining algorithm, rule mining is performed to obtain frequent manufacturing process sequences and construct a manufacturing process knowledge base.

[0008] (3) Based on the constructed manufacturing process knowledge base, a deep structured semantic model based on dual-channel attention and self-supervised learning is used as a learner to recommend and train the manufacturing process sequence, and a manufacturing process sequence recommendation model is constructed.

[0009] (4) Obtain updated part feature vectors and frequent manufacturing process sequence vectors based on the manufacturing process sequence recommendation model;

[0010] (5) Calculate the similarity between the part feature vector and the frequent manufacturing process sequence vector, and make recommendations based on the priority of the frequent manufacturing process sequences according to the similarity value.

[0011] Furthermore, the part instance model information in step (1) includes part geometric attribute features and material attribute features.

[0012] Furthermore, the relevant XML file information mentioned in step (1) includes part manufacturing process characteristics and manufacturing process sequence information.

[0013] Further, the initial manufacturing dataset described in step (1) is specifically described as follows:

[0014] IMD = {IMI} i}, i∈[1,k]

[0015] IMI i ={GF i ,MaF i ,MF i MS i}

[0016] Wherein, IMD is the initial manufacturing dataset, and IMI is the initial manufacturing dataset. i Let GF represent the i-th initial manufacturing instance in the initial manufacturing dataset, and k represent the number of instances in the initial manufacturing set. i MaF represents the geometric attribute features of the i-th instance. i Let MF represent the material property characteristics of the i-th instance. i MS represents the manufacturing process characteristics of the i-th instance. i This represents the manufacturing process sequence corresponding to the i-th instance.

[0017] Furthermore, the implementation process of step (2) is as follows:

[0018] (21) Set the minimum support threshold σ and minimum confidence threshold θ for the initial manufacturing process database MSD and sequence pattern mining indicators, as follows:

[0019] MSD = {MS i MS i ∈IMI i}, i∈[1,k]

[0020] MSD stands for Initial Manufacturing Process Database, MS i This represents the manufacturing process sequence corresponding to the i-th instance;

[0021] (22) Scan the database, count the frequency of all individual process steps, and obtain all frequent 1-itemsets F1, whose support and confidence satisfy the minimum support threshold σ, as follows:

[0022] F1={S i |support(S i )≥σ},i∈n

[0023]

[0024] Among them, S i To represent a process step, support(S) i ) represents each process step S i Support level;

[0025] (23) Based on (22), each frequent 1-itemset S i Construct its prefix projection database S Si The details are as follows:

[0026]

[0027] Among them, S Si This indicates that it includes all those starting with S. i A database of prefix projections of subsequences with prefixes;

[0028] (24) According to the prefix projection database S Si For the current prefix S i From S Si Extract all frequent itemsets S i ′, making S i S i The minimum support threshold σ and minimum confidence threshold θ are satisfied, and a new prefix projection database is constructed. The calculation method is as follows:

[0029] F Si ={S i ′∣support(S i S i ′)≥σ,confidence(S i →S i ′)≥θ}

[0030]

[0031] Among them, FSi This indicates that for the current prefix S i From S Si Extract all frequent itemsets S i ′, S SiSi ′ indicates S i S i Prefix projection database with ' as the prefix, confidence(S) i →S i ′) represents the frequently generated pattern S i S i The confidence level of ′;

[0032] (25) Continue (24) to generate new frequent pattern sequences. When the latest constructed prefix projection database is empty, stop the recursion and return the frequent manufacturing process sequence that meets the requirements. Associate the frequent manufacturing process sequence with the part instance containing the frequent sequence to construct the manufacturing process knowledge base, as follows:

[0033] PMKD = {FPS} n CFPS n}, n∈[1,m]

[0034] CFPS n ={IMI n -MS n}

[0035] Here, FPS stands for Frequent Manufacturing Process Sequence, and CFPS stands for Part Instance without Manufacturing Process Sequence, which represents the manufacturing capability of Frequent Manufacturing Process Sequence.

[0036] Furthermore, the implementation process of step (3) is as follows:

[0037] (31) Based on BERT pre-trained model, the part instance feature PF i Corresponding Frequent Manufacturing Process Sequence (FPS) i After vector transformation and embedding through all L layers, a high-dimensional vector representation is obtained. The formula for the BERT pre-trained model is as follows:

[0038] H(0) = E + P + S

[0039] H (l-1)′ =LayerNorm(H (l-1) +MultiHead_Attention(H (l-1) ))

[0040] FFN(H (l-1)′ ) = max(0, H (l-1)′ W1 (l-1) +b1 (l-1) W2 (l-1)+b2 (l-1)

[0041] H (l) =LayerNorm(H (l-1) +FFN(H (l-1)′ ))

[0042] Where E, P, and S represent word embedding, position embedding, and segment embedding of the input data, respectively, and H (l-1) The input to the (l-1)th layer of the model is represented by LayerNorm, where LayerNorm represents the layer normalization operation, MultiHead_Attention represents the multi-head attention operation, and FFN represents the feedforward neural network. W1 (l-1) W2 (l-1) Let b1 be the weight matrix of the (l-1)th layer of the model. (l-1) b2 (l-1) The bias of the (l-1)th layer;

[0043] (32) For PF i With FPS i The high-dimensional vector representation uses a dual-channel attention mechanism for vector updates, as shown in the following formula:

[0044]

[0045] Among them, z i The input vector that needs to be updated is represented by BTL, which represents the bottleneck structure layer. AvgPool and MaxPool represent average pooling and max pooling operations, respectively. σ represents the Sigmoid non-linear activation function, and BN represents batch regularization.

[0046] (33) A self-supervised learning mechanism is used to learn the representation of the updated vector. The loss function of the model is established based on joint training, as follows:

[0047] L total =L rec ({PF i FPS i})+αL self ({FPS i})

[0048] L rec =max{0,cos(PF)} i FPS j )-cos(PF i FPS i +margin}

[0049]

[0050] s(zi ,z i ′)= <z i ,z i ′i> / (||z i ||·||z i ′||)

[0051] Among them, L total Let L be the total loss function. rec Loss function for recommending manufacturing process sequences, L self Here, α is the self-supervised learning loss function, α is the regularization coefficient, margin represents the minimum distance between positive and negative sample pairs, cos represents the cosine similarity calculation, and z... i Let z represent the vector updated by the above steps for the i-th frequently manufactured sequence. i ′ represents the vector updated after feature masking for the i-th frequently generated sequence, (z i , z i ′) represents a positive sample pair, (z i , z j ′) represents a negative sample pair, τ is a temperature coefficient used to adjust the distribution of similarity scores, <z i , z i ′〉 represents z i and z i Perform the inner product, ||z i || represents z i The norm;

[0052] (34) Based on the loss value obtained in (33), the gradient of each layer is calculated, and the parameters of each layer of the network are corrected by the obtained gradient to ensure the accuracy of the algorithm.

[0053] (35) Repeat (33). When the training process reaches the upper limit of the specified number of iterations or meets the error constraint conditions, the training process ends; otherwise, training continues.

[0054] Furthermore, the implementation process of step (4) is as follows:

[0055] Based on the manufacturing process sequence recommendation model, the input features of the part to be designed and the frequent manufacturing process sequences are updated as vectors, as follows:

[0056] VPF,VFPS = RCEM(PF,FPS)

[0057] Wherein, RCEM is the manufacturing process sequence recommendation model, PF and FPS represent the features of the part to be designed and the frequent manufacturing process sequence, and VPF and VFPS represent the features of the part to be designed and the frequent manufacturing process sequence after vector update.

[0058] Furthermore, the implementation process of step (5) is as follows:

[0059] The similarity between the updated vector features of the part to be designed and the frequent manufacturing process sequences is calculated using cosine similarity, as follows:

[0060] SIM = {FPS i ,cos(VFPS i ,VPF)},i∈[1,k]

[0061] SIM represents the frequent manufacturing process sequence and its similarity to the part features; the similarity between the features of the part to be designed and each frequent manufacturing process sequence is compared and recommended according to the priority of similarity.

[0062] Furthermore, the GF i MaF i MF i The included features are:

[0063] GF i ={L i ,TH i ,OD i ,BO i ,BA i}

[0064] MaF i ={MNa i MG i MSp i ,MSt i}

[0065] MF i ={OT i WLD i PCK i CUT i SPT i}

[0066] Among them, GF i Includes length, wall thickness, outer diameter, curvature ellipticity, and curvature angle; MaF i Includes material name, material grade, material specifications, and material condition; MF i It includes oxidation treatment, welding, pickling, cutting, and painting.

[0067] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] 1. In the manufacturing process sequence recommendation method facing a large amount of complex data, this invention utilizes data mining to improve the efficiency of knowledge base construction and the learning ability of the model under skewed data distribution.

[0069] 2. Compared with existing methods, this invention not only improves the efficiency of knowledge base construction, but also considers the extraction of deep and important data features. It uses a dual-channel attention mechanism to make the model pay more attention to the implicit information of the data and improve the generalization ability of the model.

[0070] 3. This invention continues model training under skewed data distribution and combines self-supervised learning to enhance the model's representation learning on a small number of samples, ultimately achieving highly accurate manufacturing process sequence recommendation. Attached Figure Description

[0071] Figure 1 This is a flowchart of the present invention;

[0072] Figure 2 This is a flowchart of the process sequence matching algorithm based on a deep structured semantic model with dual-channel attention and self-supervised learning proposed in this invention.

[0073] Figure 3 This is a flowchart of the vector coding model based on the dual-channel attention mechanism proposed in this invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings.

[0075] like Figure 1 As shown, this invention provides a manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning. This method constructs a manufacturing process sequence recommendation model around building a knowledge base, the skewed distribution of data, and the acquisition of important information. Specifically, it includes the following steps:

[0076] Step 1: Based on the part instance model and related XML files, the instance information is read using CATIA secondary development and XPath. The part features and manufacturing process sequences from the instance model and related XML files are collected to construct an initial manufacturing dataset. The part instance model information includes geometric and material properties. The related XML file information includes manufacturing process features and manufacturing process sequences. Simultaneously, data is extracted to construct the initial manufacturing dataset, specifically described as follows:

[0077] IMD = {IMI} i}, i∈[1,k]

[0078] IMI i ={GF i ,MaF i ,MF i MS i}

[0079] Wherein, IMD is the initial manufacturing dataset, and IMI is the initial manufacturing dataset.i Let GF represent the i-th initial manufacturing instance in the initial manufacturing dataset, and k represent the number of instances in the initial manufacturing set. i MaF represents the geometric attribute features of the i-th instance. i Let MF represent the material property characteristics of the i-th instance. i MS represents the manufacturing process characteristics of the i-th instance. i This represents the manufacturing process sequence corresponding to the i-th instance. GF i MaF i MF i The included features are:

[0080] GF i ={L i ,TH i ,OD i ,BO i ,BA i}

[0081] MaF i ={MNa i MG i MSp i ,MSt i}

[0082] MF i ={OT i WLD i PCK i CUT i SPT i}

[0083] Among them, GF i Includes length (L), wall thickness (TH), outer diameter (OD), curvature ellipticity (BO), and curvature angle (BA); MaF i Includes material name (MNa), material grade (MG), material specification (MSp), and material condition (MSt); MF i It includes oxidation treatment (OT), welding (WLD), pickling (PCK), cutting (CUT), and painting (SPT).

[0084] Step 2: Perform rule mining based on the Prefixspan sequence pattern mining algorithm to obtain frequent manufacturing process sequences and build a manufacturing process knowledge base. Specifically:

[0085] The initial manufacturing process database MSD and sequence pattern mining metrics are set with minimum support threshold σ and minimum confidence threshold θ. The thresholds are selected based on multiple experiments to ensure that the mined patterns are representative but not excessive, as follows:

[0086] MSD = {MS i MS i ∈IMI i}, i∈[1,k]

[0087] MSD stands for Initial Manufacturing Process Database, MS i This represents the manufacturing process sequence corresponding to the i-th instance.

[0088] Scan the database, count the frequency of all individual process steps, and obtain all frequent 1-itemsets F1 that satisfy the minimum support threshold σ, as follows:

[0089] F1={S i |support(S i )≥σ},i∈n

[0090]

[0091] Among them, S i To represent a process step, support(S) i ) represents each process step S i Support level.

[0092] Based on each frequent 1-itemset S obtained i Construct its prefix projection database S Si The details are as follows:

[0093]

[0094] Among them, S Si This indicates that it includes all those starting with S. i A prefix projection database of subsequences with prefixes.

[0095] Based on the obtained prefix projection database S Si For the current prefix S i From S Si Extract all frequent itemsets S i ′, making S i S i The minimum support threshold σ and minimum confidence threshold θ are satisfied, and a new prefix projection database is constructed. The calculation method is as follows:

[0096] F Si ={S i ′∣support(S i S i ′)≥σ,confidence(S i →S i ′)≥θ}

[0097]

[0098] Among them, F Si This indicates that for the current prefix S i From S Si Extract all frequent itemsets S i ′, S SiSi ′ indicates S i S i Prefix projection database with ' as the prefix, confidence(S) i →S i ′) represents the frequently generated pattern S i S i The confidence level of ′.

[0099] Continue generating new frequent pattern sequences. When the latest constructed prefix projection database is empty, stop recursion and return a frequent manufacturing process sequence that meets the requirements. Associate the frequent manufacturing process sequence with the part instance containing the frequent sequence to build a manufacturing process knowledge base, as follows:

[0100] PMKD = {FPS} n CFPS n}, n∈[1,m]

[0101] CFPS n ={IMI n -MS n}

[0102] Here, FPS stands for Frequent Manufacturing Process Sequence, and CFPS stands for Part Instance without Manufacturing Process Sequence, which represents the manufacturing capability of Frequent Manufacturing Process Sequence.

[0103] Step 3: Based on the manufacturing process knowledge base built in Step 2, a deep structured semantic model based on dual-channel attention and self-supervised learning is used as the learner to train a recommendation model for manufacturing process sequences, specifically as follows: Figure 2 As shown:

[0104] Based on the manufacturing capabilities of the frequent manufacturing process sequence and the frequent manufacturing process sequence obtained in step 2, obtain the part instance feature PF. i Corresponding Frequent Manufacturing Process Sequence (FPS) i Based on the BERT pre-trained model, PF i With FPS i After vector transformation and all L layers of embedding, a high-dimensional vector representation is obtained, as shown in the following formula:

[0105] H(0) = E + P + S

[0106] H (l-1)′=LayerNorm(H (l-1) +MultiHead_Attention(H (l-1) ))

[0107] FFN(H (l-1)′ ) = max(0, H (l-1) ∪W1 (l-1) +b1 (l-1) W2 (l-1) +b2 (l-1)

[0108] H (l) =LayerNorm(H (l-1)′ +FFN(H (l-1)′ ))

[0109] Where E, P, and S represent word embedding, position embedding, and segment embedding of the input data, respectively, and H (l-1) The input to the (l-1)th layer of the model is represented by LayerNorm, where LayerNorm represents the layer normalization operation, MultiHead_Attention represents the multi-head attention operation, and FFN represents the feedforward neural network. W1 (l-1) W2 (l-1) Let b1 be the weight matrix of the (l-1)th layer of the model. (l-1) b2 (l-1) This is the bias of the (l-1)th layer.

[0110] like Figure 3 As shown, according to PF i With FPS i The high-dimensional vector representation uses a dual-channel attention mechanism for vector updates, as shown in the following formula:

[0111]

[0112] Among them, z i The input vector that needs to be updated is represented by BTL, which represents the bottleneck structure layer. AvgPool and MaxPool represent average pooling and max pooling operations, respectively. σ represents the Sigmoid non-linear activation function, and BN represents batch regularization.

[0113] A self-supervised learning method is used to learn the representation of the updated vector. The loss function of the model is established based on multi-task training, as follows:

[0114] L total =L rec ({PF i FPS i})+αL self ({FPS i})

[0115] L rec =max{0,cos(PF)} i FPS j )-cos(PF i FPS i +margin}

[0116]

[0117] s(z i ,z i ′)= <z i ,z i ′> / (||z i ||·||z i ′||)

[0118] Among them, L total Let L be the total loss function. rec Loss function for recommending manufacturing process sequences, L self Here, α is the self-supervised learning loss function, α is the regularization coefficient, margin represents the minimum distance between positive and negative sample pairs, cos represents the cosine similarity calculation, and z... i Let z represent the vector updated by the above steps for the i-th frequently manufactured sequence. i ′ represents the vector updated by the above steps for the i-th frequently manufactured sequence after feature masking, (z i , z i ′) represents a positive sample pair, (z i , z j ′) represents a negative sample pair, τ is a temperature coefficient used to adjust the distribution of similarity scores, <z i , z i ′〉 represents z i and z i Perform the inner product, ||z i || represents z i The norm of .

[0119] Based on the obtained loss value, the gradient of each layer is calculated, and the parameters of each layer are adjusted using the gradient to ensure the accuracy of the algorithm. This process is repeated until the training reaches the specified maximum number of iterations or meets the error constraint conditions; otherwise, training continues.

[0120] Step 4: Obtain the updated part feature vector and frequent manufacturing process sequence vector based on a deep structured semantic model using dual-channel attention and self-supervised learning. Specifically:

[0121] Based on the manufacturing process sequence recommendation model obtained in step 3, the input features of the part to be designed and the frequent manufacturing process sequences are updated as follows:

[0122] VPF,VFPS = RCEM(PF,FPS)

[0123] Wherein, RCEM is the manufacturing process sequence recommendation model, PF and FPS represent the features of the part to be designed and the frequent manufacturing process sequence, and VPF and VFPS represent the features of the part to be designed and the frequent manufacturing process sequence after vector update.

[0124] Step 5: Calculate the similarity between the part feature vectors obtained in Step 4 and the frequent manufacturing process sequence vectors. Based on the similarity scores, prioritize the frequent manufacturing process sequences and then make recommendations. Specifically:

[0125] The similarity between the updated vector features of the part to be designed and the frequent manufacturing process sequences is calculated using cosine similarity, as follows:

[0126] SIM = {FPS i ,cos(VFPS i ,VPF)},i∈[1,k]

[0127] Here, SIM represents the frequent manufacturing process sequence and its similarity to part features. Finally, the similarity between the features of the part to be designed and each frequent manufacturing process sequence is compared, and recommendations are made after ranking them according to the priority of similarity.

[0128] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for recommending manufacturing process sequences based on knowledge mining and self-supervised learning, characterized in that, Includes the following steps: Step (1) Based on the part instance model information and related XML file information, the instance information is read through CATIA secondary development and XPath, and the part features and manufacturing process sequences of the instance model and related XML files are collected to construct the initial manufacturing dataset; Step (2) Perform rule mining based on the Prefixspan sequence pattern mining algorithm to obtain frequent manufacturing process sequences and build a manufacturing process knowledge base; Step (3) Based on the constructed manufacturing process knowledge base, a deep structured semantic model based on dual-channel attention and self-supervised learning is used as a learner to recommend and train the manufacturing process sequence, and construct a manufacturing process sequence recommendation model. Step (4) Obtain the updated part feature vector and frequent manufacturing process sequence vector based on the manufacturing process sequence recommendation model; Step (5) Calculate the similarity between the part feature vector and the frequent manufacturing process sequence vector, and then recommend the frequent manufacturing process sequences based on the priority of the similarity values. The initial manufacturing dataset mentioned in step (1) is specifically described as follows: ； ； Wherein, IMD is the initial manufacturing dataset, and IMI is the initial manufacturing dataset. i Let GF represent the i-th initial manufacturing instance in the initial manufacturing dataset, and k represent the number of instances in the initial manufacturing set. i MaF represents the geometric attribute features of the i-th instance. i Let MF represent the material property characteristics of the i-th instance. i Indicates the first Manufacturing process characteristics of an instance, MS i This represents the manufacturing process sequence corresponding to the i-th instance; The implementation process of step (2) is as follows: (21) Set the minimum support threshold σ and minimum confidence threshold θ for the initial manufacturing process database MSD and sequence pattern mining indicators, as follows: ； MSD stands for Initial Manufacturing Process Database, MS i This represents the manufacturing process sequence corresponding to the i-th instance; (22) Scan the database, count the frequency of all individual process steps, and obtain all frequent 1-itemsets F1, whose support and confidence satisfy the minimum support threshold σ, as follows: ； ； Among them, S i To represent a process step, support(S) i ) represents each process step S i Support level; (23) Based on (22), each frequent 1-itemset S i Construct its prefix projection database S Si The details are as follows: ； Among them, S Si This indicates that it includes all those starting with S. i A database of prefix projections of subsequences with prefixes; (24) According to the prefix projection database S Si For the current prefix S i From S Si Extract all frequent itemsets S i ′, making S i S i The minimum support threshold σ and minimum confidence threshold θ are satisfied, and a new prefix projection database is constructed. The calculation method is as follows: ； ； ； Among them, F Si This indicates that for the current prefix S i From S Si Extract all frequent itemsets S i ′, S SiSi′ Indicated by S i S i Prefix projection database with ' as the prefix, confidence(S) i →S i ′) represents the frequently generated pattern S i S i The confidence level of ′; (25) Continue (24) to generate new frequent pattern sequences. When the latest constructed prefix projection database is empty, stop the recursion and return the frequent manufacturing process sequence that meets the requirements. Associate the frequent manufacturing process sequence with the part instance containing the frequent manufacturing process sequence to construct the manufacturing process knowledge base, as follows: ； ； Here, FPS stands for Frequent Manufacturing Process Sequence, and CFPS stands for Part Instance without Manufacturing Process Sequence, which represents the manufacturing capability of Frequent Manufacturing Process Sequence.

2. The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning according to claim 1, characterized in that, The part instance model information in step (1) includes part geometric property features and material property features.

3. The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning according to claim 1, characterized in that, The relevant XML file information mentioned in step (1) includes part manufacturing process characteristics and manufacturing process sequence information.

4. The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning according to claim 1, characterized in that, The implementation process of step (3) is as follows: (31) PF of part instance features based on BERT pre-trained model i Corresponding Frequent Manufacturing Process Sequence (FPS) i After vector transformation and embedding through all L layers, a high-dimensional vector representation is obtained. The formula for the BERT pre-trained model is as follows: ； ； ； ； Where E, P, and S represent word embedding, position embedding, and segment embedding of the input data, respectively, and H (l-1) The input to layer l-1 of the model is represented by `LayerNorm`, the layer normalization operation is represented by `MulitiHead_Attention`, and `FFN` represents a feedforward neural network. , This is the weight matrix of the (l-1)th layer of the model. , The bias of the (l-1)th layer; (32) For PF i With FPS i The high-dimensional vector representation uses a dual-channel attention mechanism for vector updates, as shown in the following formula: ； ； Among them, z i This represents the input vector that needs to be updated. BTL represents the bottleneck structure layer. AvgPool and MaxPool represent the average pooling operation and the max pooling operation, respectively. represents the Sigmoid non-linear activation function, and BN represents batch regularization; (33) A self-supervised learning mechanism is used to learn the representation of the updated vector. The loss function of the model is established based on joint training, as follows: ； ； ； ； Among them, L total Let L be the total loss function. rec Loss function for recommending manufacturing process sequences, L self For self-supervised learning loss function, Here, is the regularization coefficient, margin represents the minimum distance between positive and negative sample pairs, cos represents the cosine similarity calculation, and z... i Let z represent the vector updated by the above steps for the i-th frequently manufactured sequence. i ' represents the vector updated after feature masking for the i-th frequently generated sequence, (z) i , z i ′) represents a positive sample pair, (z) i , z j ′) represents a negative sample pair. It is a temperature coefficient used to adjust the distribution of similarity scores, <z i , z i ′〉 represents z i and z i Perform the inner product, ||z i || represents z i The norm; (34) Based on the loss value obtained in (33), the gradient of each layer is calculated, and the parameters of each layer of the network are corrected by the obtained gradient to ensure the accuracy of the algorithm; (35) Repeat (33). When the training process reaches the upper limit of the specified number of iterations or meets the error constraint conditions, the training process ends; otherwise, training continues.

5. The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning according to claim 4, characterized in that, The implementation process of step (4) is as follows: Based on the manufacturing process sequence recommendation model, the input features of the part to be designed and the frequent manufacturing process sequences are updated as vectors, as follows: ； Wherein, RCEM is the manufacturing process sequence recommendation model, PF and FPS represent the features of the part to be designed instance and the frequent manufacturing process sequence, and VPF and VFPS represent the features of the part to be designed and the frequent manufacturing process sequence after vector update.

6. The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning according to claim 5, characterized in that, The implementation process of step (5) is as follows: The similarity between the updated vector features of the part to be designed and the frequent manufacturing process sequences is calculated using cosine similarity, as follows: ； SIM represents frequent manufacturing process sequences and their similarity to part features; the similarity between the features of the part to be designed and each frequent manufacturing process sequence is compared, and recommendations are made after ranking them according to the priority of similarity.

7. The manufacturing process sequence recommendation method based on knowledge mining and self-supervised learning according to claim 1, characterized in that, The GF i MaF i MF i The included features are: ； ； ； Among them, GF i Includes length Wall thickness , outer diameter Curvature ellipticity Bending angle ;MaF i Includes material name Material grade Material specifications Material condition ;MF i Includes oxidation treatment ,welding Pickling Cutting Spray painting .

Images